Import AI

Import AI 155: Mastering robots with the ‘DRIVE’ dataset; facial recognition for monkeys; and why AI development is a collective action problem.

Chinese company seeks smarter robots with ‘DRIVE’ dataset:
…Crowds? Check. Trashcan-sized robots? Check. A challenging navigation and mapping benchmark? Check…
Researchers with Chinese robot company Segway Robotics Inc have developed the ‘DRIVE’ dataset and benchmark, which is designed to help researchers develop smarter delivery robots. 

   The company did the research because it wants to encourage research in an area relevant to its business, and because of larger macroeconomic trends: “The online shopping and on-demand food delivery market in China has been growing at a rate of 30%-50% per year, leading to labor shotage and rising delivery cost,” the researchers write. “Delivery robots have the potential to solve the dilemma caused by the growing consumer demand and decreasing delivery workforce.” 

Robots! Each Segway robot used to gather the dataset is equipped with a RealSense visual inertial sensor, two wheel encoders, and a Hokuyuo 2D lidar. 

The DRIVE dataset: The dataset consists of 100 movement sequences across five different indoor locations, and was collected by robots over the course of one year. It is designed to be extremely challenging, and incorporates the following confounding factors and traits:

  • Commodity, aka cheap, inertial measurement units

  • Busy: The gathered data includes scenes with many moving people and objects, which can break brittle AI systems

  • Similar, similar: Some of the environments are superficially similar to eachother, which could trigger misclassification. Additionally, some of the places in the environments lack texture or include numerous reflections and shadows, making it harder for robots to visually analyze their environment. Additionally, some of the environments have bumpy or rough surfaces.

  • Hurry up and wait: Some of the datasets include long sequences in which the robot is stationery (which makes it difficult to estimate depth), while at other times the robots perform rapid rotations (which can lead to motion blur and wheels slipping on the ground). 

Why this matters: Datasets unlock AI progress, letting large numbers of people work together on shared challenges. Additionally, the creation of datasets usually imply specific business and research priorities, so the arrival of things like the DRIVE Benchmark point to broader maturation in smart, mobile robots.
   Read more: Segway DRIVE Benchmark: Place Recognition and SLAM Data Collected by A Fleet of Delivery Robots (Arxiv).
   Find out more about the benchmark here (Segway DRIVE website).

####################################################

You’ve heard of face identification. What about Primate face identification?
…Towards a future where we automatically scan and surveil the world around us…
Researchers with the Indraprastha Institute of Information Technology Delhi and the Wildlife Institute of India have teamed up to develop a system capable of identifying monkeys in the wild and have linked this to a crowd-sourced app, letting the “general public, professional monkey catchers and field biologists” crowd source images of monkeys for training larger, smarter models. 

Why do this? Monkeys are a bit of a nuisance in Indian urban and semi-urban environments, the researchers write, so have designed the system to use data captured ‘in the wild’, helping people build systems to surveil and analyze primates in challenging contexts. “Typically, we expect the images to be captured in uncontrolled outdoor scenarios, leading to significant variations in facial pose and lighting”. 

Datasets: 

  • Rhesus Macaque Dataset: 7679 images / 93 individuals. 
  • Chimpanzee Dataset: 7166 images / 90 primates. Pictures span good quality images from a Zoo, as well as uncontrolled images from a national park.

Results: The system outperforms a variety of baselines and sets a new state of the art across four validation scores, typically via a greater than 2 point absolute increase in performance, and sometimes via as much as a 6 or greater point increase. Their system is trained with a couple of different loss functions designed to capture smaller geometric features across faces, making the model more robust across multiple data distributions. 

Why this matters: This research is an indication of how as AI has matured we’ve started to see it being used as a kind of general-purpose utility, with researchers mixing and matching different techniques and datasets, making slight tweaks, and solving tasks for socially relevant applications. It’s particularly interesting to see this approach integrated with a crowd sourced app, pointing to a future where populations are able to collaboratively measure, analyze, and quantify the world around them.
   Read more: Primate Face Identification in the Wild (Arxiv)

####################################################

What Recursion’s big dataset release means for drug discovery:
…RxRx1 dataset designed to encourage “machine learning on large biological datasets to impact drug discovery and development”…
Recursion Pharmaceuticals, a company that uses AI for drug discovery, has released RxRx1, a 296GB dataset consisting of 125,510 images across 1,108 classes; an ImageNet-scale dataset, except instead of containing pictures of cats and dogs it contains pictures of human cells, to help scientists train AI systems to observe patterns across them, and generate insights for drug development. 

The challenges of biology: Biological datasets can be challenging for image recognition algorithms due to variation across cell samples, and other factors present during data sampling, such as temperature, humidity, reagent concentration and so on. RxRx1 contains data from 51 instances of the same experiment, which should help scientists develop algorithms that are robust to the changes across experiments, and are thus able to learn underlying patterns in the data.

What parts of AI research could RxRx1 help with? Recursion has three main ideas:

  • Generalization: The dataset is useful for refining techniques like transfer learning and domain adaptation.
  • Context Modeling: Each RxRx1 image ships with a detailed metadata, so researchers can experiment with this as an additional form of signal. 
  • Computer Vision: RxRx1 “presents a very different data distribution than is found in most publicly available imaging datasets,” Recursion writes. “These differences include the relative independence of many of the channels (unlike RGB images) and the fact that each example is one of a population of objects treated similarly as opposed to singletons.” 

Why this matters: We’re entering an era where people will start to employ large-scale machine learning to revolutionize medicine; tracking usage of datasets like RxRx1 and the results of a planned NeurIPS 2019 competition will help give us a sense of progress here and what it might mean for medicine and drug design.
   Read more: RxRx1 official website (RxRx.ai).

####################################################

Why AI could leave people with disabilities behind:
…Think bias is a problem now? Wait until systems are deployed more widely…
Researchers with Microsoft and the Human-Computer Interaction Institute at Carnegie Mellon University have outlined how people with disabilities could be left behind by AI advances. People with disabilities could have trouble accessing the benefits of AI systems due to issues of fairness and bias inherent to machine learning, according to a position paper from researchers with Microsoft and the Human-Computer Interaction Institute at Carnegie Mellon University. To deal with some of these issues, they propose a research agenda to help remedy these shortcomings in AI systems. The agenda contains four key activities:

  • Identify ways in which inclusion issues for people with disabilities could impact AI systems
  • Test inclusion hypotheses to understand failure scenarios
  • Create benchmark datasets to support replication and inclusion
  • Develop new modeling, bias mitigation, and error measurement techniques 

It’s all about representation: So, how might we expect AI systems to fail for people with disabilities? The authors survey current systems and provide some ideas. Spoiler alert: Mostly, these systems will fail to work for people with disabilities because they will have been designed by people who are neither disabled, nor are educated about the needs of people with disabilities.

  • Computer Vision: It’s likely that facial recognition will not work well for people with differences in facial features and expressions (eg, people with Down’s syndrome) not anticipated by system designers; face recognition could also not work for blind people, who may have differences in eye anatomy or be wearing medical or cosmetic aids. For similar reasons, we can expect systems designed to recognize certain bodytypes failing for some people. Additionally, object/scene/text recognition systems are likely to break more frequently for poorly sighted people, as the pictures poorly sighted people take are very different to those taken by sighed people. 
  • Speech Systems: Speech recognition systems won’t work for people that have speech disabilities; we may also need more granular metrics beyond things like Word Error Rate to best model how well systems work for different people. Similarly, speaker analysis systems will need to be trained with different datasets to accurately hear people with disabilities. 
  • Text Analysis: These systems will need to be designed to correct for errors that emerge under certain disabilities (for instance, dyslexia), and will need to account for people that write in different emotional registers to typical people. 

Why this matters: AI is an accelerant and a magnifier of whatever context it is deployed in due to the scale at which it operates, the number of automatic judgements it makes, and the increasingly comprehensive deployment of Ai-based techniques across society. Therefore, if we don’t think very carefully about how AI may or may not ‘see’ or ‘understand’ certain types of people, we could harm people or cut them off from accessing its benefits. (On the – extremely minor – plus side, this research suggests that people with disabilities may be harder to surveil than other people, for now.)
   Read more: Toward Fairness in AI for People with Disabilities: A Research Roadmap (Arxiv)

####################################################

‘Visus’ software provides quality assurance for model training:
…Now that models can design themselves, we need software to manage this…
Researchers with New York University have developed Visus, software that makes it easier for people to build models, evolve models, and manage the associated data processing pipelines needed to train them. It’s a tool that represents the broader industrialization of the AI community, and prefigures larger uses of ML across society.

What is Visus? The software gives AI developers a software interface that lets them define a problem, explore summaries of the input dataset, augment the data, and then explore and compare different models according to their performance scores and prediction outputs. The software is presented via a nicely designed user interface, making it more approachable than tools solely accessible via the command line. 

What can it do? What can’t it do! Visus is ‘kitchen sink software’, in the sense that it contains a vast number of features for tasks like exploratory data analysis, problem specification, data augmentation, model generation and selection, and confirmatory data analysis, and so on. 

Example use case: The researchers outline a hypothetical example where the New York City Department of Transportation uses Visus to figure out policies that it can enact which can reduce traffic fatalities. Here, they’d use Visus first to analyze the dataset about traffic collisions, then can select a variable in the dataset (for instance, number of collisions) that they’d want to predict, then ask Visus to perform a model search (otherwise known as ‘AutoML’), where it tries to find appropriate machine learning models to use to achieve the objective. Once it comes up with models, the user can also try to augment the underlying dataset, and then iterate on model design and selection again. 

Why this matters: Systems like ‘Visus’ are part of the industrialization of AI, as they take a bunch of incredibly complicated things like data augmentation and model design and analysis, then port it into more user-friendly software packages that broaden the number of people able to use such systems. This is like shifting away from artisanal individualized production to repeatable, system-based production. The outcome of adoption of tools like Visus will be more people using more AI systems across society – which will further change society.
   Read more: Visus: An Interactive System for Automatic Machine Learning Model Building and Curation (Arxiv).

####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

Collective action problems for safe AI:
In many industries, profit-seeking firms are incentivised to invest in product safety. This is generally because they have internalised the costs of safety failures via regulation, liability, and consumer behaviour. Consider the cost to a car manufacturer of a critical safety failure – they will have to recall the product, they will be liable to fines and litigation, and they will suffer reputational damage. AI firms are subject to these incentives, but they appear to be weaker. Their products are difficult for manufacturers, consumers, and regulators to assess for safety; it is difficult to construct effective regulation; and many of the potential harms might be hard to internalise.

Competition: Another special feature about AI development is the possibility of discontinuous and/or very rapid progress. If firms believe this, they likely believe that there are significant payoffs to the first firm to make a particular breakthrough or to ‘pull ahead’ from competitors. This increases the costs of investing in safety, by increasing the expected benefits of faster development. This assumption may not hold true, which would make the situation more benign, but it is important to consider what this ‘worst-case’ for responsible development.

Cooperation: A simple model of this problem is a two-player game, where two firms face a decision to cooperate (maintain some level of investment in safety) or defect (fail to maintain this level). This allows us to see factors that can increase the likelihood of cooperation, by making it rational for each firm to do so: high trust that others will cooperate; shared upside from mutual cooperation; shared downside from mutual defection; smaller benefits to not reciprocating cooperation; and lower costs to unreciprocated cooperation.

Four strategies: This analysis can help identify strategies for increasing cooperation on responsible development: dispelling incorrect beliefs about responsible AI development; promoting inter-firm collaboration on projects; opening AI development to appropriate oversight and feedback; and creating stronger incentives to safe practices.
   Read more: The Role of Cooperation in Responsible AI Development (arXiv).
   Read more: Why Responsible AI Development Needs Cooperation on Safety (OpenAI Blog).
   Further thoughts on the project from corresponding author Dr Amanda Askell (Twitter).

####################################################

Tech Tales: 

Stop-Start Computing

And so after the Climate Accords and the Generational Crime Rulings and the loss of some 20% of the world’s land surface to a combination of heat and/or flooding, after all of this society carried on. We were hot and we were sick and there were way too many of us, but we carried on. 

We kept on moving little and big chunks of mass around on planet earth, and as we moved this stuff we mixed up the atmosphere and the underground and we changed our air and made it worse, but we carried on. 

And all through this we used our computers. We used our phones to watch old movies of ‘the times before’. We listened to music from prior decades. We played games in which the planet was covered in forests, or where we were neanderthals playing with axes in a kind of wilderness, or ones where we rode out into space and managed vast interstellar armies. Our simulations and our software and our entertainment got better and better and so we used it more and more, and we carried on. 

Everything has a breaking point. At some point computers started using so much energy that even with central planning and the imposition of controls, electrical utilities couldn’t keep up. Thirty percent of the electricity in some countries went to computers. In some smaller countries based around high-tech services, it was even higher. Data centers found themselves periodically running on backup generators – old salvaged WW2 diesel engines from submarines – and sometimes the power ran out entirely and these big computer cathedrals stood idle, mute blocks surrounded by farmland or forest or high-altitude steppes and deserts. 

So after we hit our limit we created the coins as part of the Centrally Managed Sustainable Compute Initiative. We were meant to call them ‘compute tokens’ but everyone called them coins, and we were meant to call the computation power we exchanged these coins for the Shared Societal Computer but everyone just called it the timeshare. 

So now here’s how it works: 

  • If you’re poor, you use a coin and you access lumps of computation and storage, rationed out according to the complex interplay of heat and consumption and climate. 
  • If you’re rich, you spend extra for Premium Compute Credits. 
  • If you’re ultrarich, you build yourself a powerplant or better yet something renewable – geothermal or wind or solar. Then you build your facility and you use that computation for yourself. 

Private data centers will be outlawed soon, people say. There’s talk of using all of the compute left in the world to save the world – something about simulating the impossible complexity of the earth, and finding a way to carry on. 

Things that inspired this story: The energy consumption of Bitcoin and large-scale AI models; climate change; inevitability.

Import AI 154: Teaching computers how to plan; DeepNude is where dual-use meets pornography; and what happens when we test machine translation systems on real-world data

Can computers learn to plan? Stanford researchers thinks so:
…Turns out being able to plan is similar to figuring out where you are and where you’ve been…
Researchers with Stanford University have developed a system that can watch instructional videos on YouTube and learn to look at the start and end of a new video then figure out the appropriate order of actions to take to transition from beginning to end.

What’s so hard about this? The real world involves such a vast combinatorial set of possibilities that traditional planning approaches (mostly) aren’t able to scale to work within it. “One can imagine an indefinitely growing semantic state space, which prevents the application of classical symbolic planning approaches that require a given set of predicates for a well-defined state space”. To get around this, they instead try to learn everything in a latent space, essentially slurping in reality and turning it into features, which they then use to map actions and observations into sequences, helping them figure out a plan.

Two models to learn the latent space:
   The system that derives the latent space and the transformations within it has two main components:

  • A transition model, which predicts the next state based on the current state and action.
  • A conjugate constraint model which maps current actions to past actions.

   The full model takes in a video and essentially learns the transitions between states by sliding these two models along through time to the desire goal state, sampling actions and then learns the next state. 

Two approaches to planning: The researchers experiment with two planning approaches, both of which rely on the features mined by the main system. One approach tries to map current and goal observations into a latent space while also mapping actions to prior actions, then samples from different actions to use to solve its task. The other approach is called ‘walkthrough planning’ and outputs the visual observations between the current and goal state; this is a less direct approach as it doesn’t output actions, but could serve as a useful reward signal for another system. 

Dataset: For this work, they use the CrossTask instructional video dataset, which is a compilation of videos showing 83 different tasks, involving things like grilling steak, making pancakes, changing a tire, and so on.

Testing: Spoiler alert – this kind of task is extremely hard, so get ready for some stay-in-your-chair results. In tests, the researchers find their system using the traditional planning approach can obtain accuracies of around 31.29% tests, with an overall success rate of 12.18%. This compares to a prior state-of-the-art of 24.39% accuracy and 2.89% success rate for ‘Universal Planning Networks’ (Import AI #90). (Note: UPN is the closest thing to compare to, but has some subtle differences making a direct comparison difficult). They show that the same system when using walkthrough planning can significantly improve scores over prior state-of-the-art systems as well – “our full model is able to plan the correct order for all video clips”, they write, compared to baselines which typically fail. 

Why this matters: We’re starting to see AI systems that use the big, learnable engines used in deep learning research as part of more deliberately structured systems to tackle specific tasks, like learning transitions and plans for video walkthroughs. Planning is an essential part of AI, and being able to learn plans and disentangle plans from actions (and learn appropriate associations) is an inherently complex task; progress here can give us a better sense for progress in the field of AI
   Read more: Procedure Planning in Instructional Videos (Arxiv)

####################################################

DeepNude: Dual Use concerns meet Pornography; trouble ensues:
…Rock, meet hard place…
What would a person look like without their clothes? That’s something people can imagine fairly easily, but has been difficult for AI systems. That is, until we developed a whole bunch of recent systems capable of modeling data distributions and generating synthetic versions of said data; these techniques contributed to the rise of things like ‘deepfakes’ which let people superimpose the face of one person on that of another in a video. Recently, someone took this a step further with a software tool called DeepNude which automatically removes the clothes of (predominantly women), rendering synthetic images of them in the nude. 

Blowback, phase one: The initial DeepNude blowback centered on the dubious motivation for the project and the immense likelihood of the software being used to troll, harass, and abuse women. Coverage in Vice led to such outcry from the community that the creator of DeepNude took the application down – but not before others had implemented the same capabilities in other software and distributed it around the web. 

Rapid proliferation makes norms difficult: Just a couple of days after taking the app down, the creator posted the code of the application to GitHub, saying that because the DeepNude application had already been replicated widely, there was no purpose in keeping the original code private, so they published it online. 

Why this matters: DeepNude is an illustration of the larger issues inherent to increasingly powerful AI systems; these things have got really powerful and can be used in a variety of different applications and are also, perhaps unintuitively, relatively easy to program and put together once you have some pre-trained networks lying around (and the norms of publication mean this is always the case). How we figure out new norms around development and publication of such technology will have a significant influence on what happens in society, and if we’re not careful we could enable more things like DeepNude.
   Read the statement justifying code release: Official DeepNude Algorithm (DeepNude GitHub).
   Read more: This Horrifying App Undresses a Photo of any Woman With a Single Click (Vice). (A special ImportAI shoutout to Samantha Cole, the journalist behind this story; Samantha was the first journalist to cover deepfakes back in 2017 and has been on this beat doing detailed work for a while. Worth a follow!)

####################################################

Have no pity for robots? Watch these self-driving cars try to tackle San Francisco:
A short video from Cruise, a self-driving car service owned by General Motors, shows how its cars can now deal with double-parked cars in San Francisco, California.
    Check out the video here (official Cruise Twitter).

####################################################

Think AI services are consistent across cloud providers? Think again:
…Study identifies significant differences in AI inferences made by Google, Amazon, and Microsoft…
Different AI cloud providers have different capabilities, and these under-documented differences could cause problems for software developers, according to research from computer science researchers with Deakin University and Monash University in Australia. In a study, they explore the differences between image labeling AI services from Amazon (“AWS Rekognition”), Google (“Google Cloud Vision”) and Microsoft (“Azure Computer Vision”). The researchers try to work out if “computer vision services, as they currently stand, offer consistent behavior, and if not, how is this conveyed to developers (if it is at all)?”

Developers may not realize that services can vary from cloud provider to provider, the researchers write; this is because if you look at the underlying storage and compute systems across major cloud providers like Microsoft or Amazon or Google you find that they’re very comparable, whereas differences in the quality of AI services are much less easy to work out from product descriptions. (For instance, one basic example is the labels services output when classifying objects; one service may describe a dog as both a ‘collie’ and a ‘border collie’, while another may use just one (or none) of these labels, etc.) 

Datasets and study length: The authors used three datasets to evaluate the services; two self-developed ones – a small one containing 30 images and a large one containing 1,650 ones, and a public dataset called COCOVal17, which contains 5,000 images. The study took place over 11 months and had two main experimental phases: a 13 week period from April to August 2018 and a 17 week period from November 2018 to March 2019. 

Methodology: They test the cloud services for six traits: the consistency of the top label assigned to an image from each service; the ‘semantic consistency’ of multiple labels returned by the same service; the confidence level of each service’s top label prediction; the consistency of these confidence intervals across multiple services; the consistency of the top label over time (aka, does it change); and the consistency of the top label’s confidence over time. 

Three main discoveries: The paper generates evidence for three concerning traits in clouds, which are:

Computer vision services do not respond with consistent outputs between services, given the same input image. 

  • Outputs from computer vision services are non-deterministic and evolving, and the same service can change its top-most response over time given the same input image. 
  • Computer vision services do not effectively communicate this evolution and instability, introducing risk into engineering these systems. 

Why this matters: Commercial AI systems can be non-repeatable and non-reliable, and this study shows that multiple AI systems developed by different providers can be even more inconsistent with one another over time. This is going to be a challenging issue, as it makes it easier for developers to get ‘locked in’ to the specific capabilities of a single service, and also makes application portability difficult. Additionally, these issues will make it harder for people to build AI services that are composed out of multiple distinct AI services from different clouds, as these systems will not have predictable performance capabilities.
   Read more: Losing Confidence in Quality: Unspoken Evolution of Computer Vision Services (Arxiv).

####################################################

Stealing people’s skeletons with deep learning:
…XNect lets researchers do real-time multi-person pose estimation via a single RGB camera…
How do you use a single camera to track multiple people and their pose as they move around? That’s a question being worked on by researchers with the Max Planck Institute for Informatics, EPFL, and the University of Saarland. The try to solve this problem via a neural network architecture that encodes and decodes poses of people, which is also implemented efficiently enough to run in real-time from a single camera feed. The system uses two networks; one which focuses on learning to reason about individual body joints, and another which tries to jointly reason about all body joints. 

Special components for better performance: Like some bits of AI research, this work takes a bunch of known-good stuff, and then pushes it forward on a task-specific dimension. Here, they develop a convolutional neural network architecture called SelecSLS Net, which “employs selective long and short range concatenation-skip connections to promote information flow across network layers which allows to use fewer features leading to a much faster inference time but comparable accuracy in comparison to ResNet-50”. 

Real-time performance: Most of the work here has involved increasing the efficiency of the system so it can process footage from video cameras in real-time (when running on an NVIDIA GTX 1080Ti and a Xeon E5). In terms of performance, the system marginally outperforms a more standard system that uses a typical residual network, while being far more efficient when it comes to runtime. 

Why this matters: It’s becoming trivial for computers to look at people, model each of them as a wireframe skeleton, and then compute over that. This is a classic omni-use capability; we could imagine such a system being used to automatically port people into simulated virtual worlds, or to plug them into a large-scale surveillance system to analyze their body movements and characterize the behavior of the crowd. How society deals with the challenges of such a multi-purpose technology remain to be seen.
   Read more: XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera (Arxiv).

####################################################

Think network design is hard? Try it where every network point is a drone:
…Researchers show how to build dynamic networks out of patrolling drones…
Researchers with Alpen-Adria-Universitat Klagenfurt, Austria, have developed “a novel collaborative data delivery approach where UAVs transport data in a store-and-forward fashion”. What this means is they develop a system that automatically plans the flight paths of fleets of drones so that the drones at the front of the formation periodically overlap in communication range with UAVs behind them, which then overlap in communication range with other, even more distant UAVs. The essential idea behind the research is to use fast drone-to-drone communications systems to hoover up data via exploration drones at the limits of a formation, then squirt this data back to a base station via the drones themselves. The next step for the research is to use “more sophisticated scheduling of UAVs to minimize the number of idle UAVs (that do neither sensing nor transporting data) at each time step”. 

Why this matters: Drones are going to let people form ad-hoc computation and storage systems, and approaches like this suggest the shape of numerous ‘flying internets’ that we could imagine in the future.
   Read more: Persistent Multi-UAV Surveillance with Data Latency Constraints (Arxiv).

####################################################

Pushing machine translation systems to the limit with real, messy data:
…Machine translation robustness competition shows what it takes to work in the real world…
Researchers from Facebook AI Research, Carnegie Mellon University, Harvard University, MIT, the Qatar Computing Research Institute, Google, and Johns Hopkins University, have published the results of the “first shared task on machine translation robustness”. The goal of this task is to give people better intuitions about how well machine translation models deal with “orthographic variations, grammatical errors, and other linguistic phenomena common in user-generated content”. 

Competitions, what are they good for? The researchers hope that systems which do well at this task will use better modelling, training and adaptation techniques, or may learn from large amounts of unlabeled data. And indeed, entered systems did use a variety of additional techniques to increase their performance, such as data cleaning, data augmentation, fine-tuning, ensembles of models, and more. 

Datasets: The datasets were “collected from Reddit, filtered out for noisy comments using a sub-word language modeling criterion and translated by professional translators”

Results: As this competition explores robustness in the context of a competition, it’s perhaps less meaningful to focus on the quantitative results, and instead discuss the trends seen among the entries. Some of the main things seen by the competition organizers are: stronger submissions were typically stronger across the board; out-of-domain generalization is important (so having systems that can deal with words they haven’t seen before); being able to accurately model upper and lower case text, as well as the use of special characters, is useful; it can be difficult to learn to translate sentences written in slang, 

Why this matters: Competitions like this give us a better sense of the real-world progress of AI systems, helping us understand what it takes to build systems that work over real data, as opposed to highly-constrained or specifically structured test sets.
   Read more: Findings of the FIrst Shared Task on Machine Translation Robustness (Arxiv).

####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

Axon Ethics Board— no face recognition on police body cameras:
Axon, who make technologies for law enforcement, established an AI Ethics Board back in 2018 to look at the ethical implications of their products. The board has just released their first report, looking at ethical issues surrounding face recognition, particularly on police body cameras—Axon’s core product.

The board: Axon was an early mover in establishing an AI ethics board. The board’s members are drawn from law enforcement, civil rights groups, policy, academia, and tech. Among the lessons learned, the Board emphasizes the importance of board involvement at an early stage in product development (ideally before the design stage), so that they can suggest changes before they become too costly for the company.

Six major conclusions:
  (1) Face recognition technology is currently not reliable enough to justify use on body-cameras. Far greater accuracy, and equal performance across different populations are needed before deployment.
  (2) In assessing face recognition algorithms, it is important to separate false positive and false negative rates. There are real trade offs between the two, which depend on use cases. E.g. in identifying a missing person, more false positives might be cost worth bearing to minimize false negatives. Whereas in enforcement scenarios, it might be more important to minimize false positives, due to the potential harms from police interacting with innocent people on mistaken information.
  (3) The Board does not endorse the development of face recognition technology that can be completely customised by users, to prevent misuse. This requires technological controls by product manufacturers, but will increasingly also require government regulation.
  (4) No jurisdiction should adopt the technology without going through transparent, democratic processes. At present, big decisions affecting the public are being made by law enforcement alone, e.g. whether to include drivers license photos in face databases.
  (5) Development of products should be premised on evidence-based (and not merely theoretical) benefits.
  (6) When assessing costs and benefits of potential use cases, one must take into account the realities of policing in particular jurisdictions, and technological limitations.
  Read more: First Report of the Axon AI & Policing Ethics Board (Axon).
  Read more: Press release (Axon).

####################################################

NIST releases plan on AI standards:
The White House’s executive order on AI, released in February, included an instruction for NIST to make “a plan for Federal engagement in the development of technical standards and related tools in support of reliable, robust, and trustworthy systems that use AI technologies.” NIST have released a draft plan, and are accepting public input until July 19, before delivering a final document in August. Recommendations: NIST recommends that the government “bolster AI standards-related knowledge, leadership, and coordination among federal agencies; promote focused research on the ‘trustworthiness’ of AI; support and expand public-private partnerships; and engage with international parties.”

Why it matters: The US is keen to lead international efforts in standards-setting. Historically, international standards have governed policy externalities in cybersecurity, sustainability, and safety. Given the challenges of trust and coordinating safe practices in AI development and deployment, standards setting could play an important role.
  Read more: U.S. Leadership in AI: a Plan for Federal Engagement in Developing Technical Standards and Related Tools (NIST).

####################################################

Tech tales

Dreamworld versus Reality versus Government

After the traceable content accords were enacted people changed how they approached themselves – nude photos aren’t so fun if you know your camera is cryptographically signing them and tying them to you then uploading that information to some vast database hosted by a company or a state. 

The same thing happened for a lot of memes and meme-fodder: it’s not obviously a good idea to record yourself downing ten beers on an amusement park ride if you’re subsequently going to pursue a career in politics, nor does it seem like a smart thing to participate in overtly political pranks if you think you might pursue a career in law enforcement. 

The internet got… quiet? It was still full of noise and commotion and discussion, but the edge had been taken off a little. Of course, when we lost the edge we lost a lot of pain: it’s harder to produce terrorist content if it is traced back to your phone or camera or whatever, and it’s harder for other people to fake as much of it when it stops being, as they say, a ‘desirable media target’.

It didn’t take long for people to figure out a work around: artificial intelligence. Specifically, using large generative models to create images and, later, audio, and even later after that, videos, which could synthesize the things they wanted to create or record, but couldn’t send or do anymore. Teens started sending eachother impressionistic, smeared videos of teen-like creatures doing teen-like pranks. Someone invented some software called U.S.A which stood for Universal Sex Avatar and teens started sending eachother ‘AIelfies’ (pronounced elfeez) which showed nude-like human-like things doing sexual-like stuff. Even the terrorists got involved and started pumping out propaganda that was procedural and generative. 

Now the internet has two layers: the reality-layer and what people have taken to calling the dreamworld. In the reality-layer things are ever-more controlled and people conduct themselves knowing that what they do will be knowable and identifiable most-likely forever; everyone’s a politician, essentially. In the dreamworld, people experiment with themselves, and everyone has a few illicit channels on their messaging apps through which they let people send them dreamworld content, and through which they can anonymously and non-anonymously send their own visions into the world. 

The intelligence agencies are trying to learn about the dreamworld, people say. Knowing the difference between what known individuals publicly present and what the ghostly mass of civilization illicitly sends to itself is a valuable thing, say certain sour-faced people who are responsible for terrible tools that ward off against more terrible things. “The difference between presented self and imagined self is where identity resides,” says one of them in a no-phone presentation to other sour-faced people. “If we can learn how society chooses to separate the two, perhaps we can identify the character of our society. If we can do that, we can change the character.”

And so the terrible slow engines are working now, chewing through our dreamworld, invisible to us, but us increasingly aware of them. Where shall we go next, we wonder? What manifestation shall our individuality take next?

Things that inspired this story: Generative adversarial networks; DeepNude; DeepFakes; underground communities; private messaging infrastructures; the conversation of all of physical reality into digital simulacra.

 

Import AI 153: Why not all cloud AI services are created equally; making more repeatable robots with PyRep; and surveying crops with drones

Chinese scientists set new state-of-the-art in crowd counting:
…The secret? Dense dilated convolutions with residual connections…
Researchers with the Chinese Academy of Sciences and the University of Science and Technology of China in Hefei have developed a new system for counting the number of people in a crowd. The system, called DSNet, sets state-of-the-art performance on four significant datasets, and should serve as a reminder that AI is an omni-use technology, where progression on fundamental techniques (eg: residual networks) can directly translate to advances in tools for surveillance. 

Dense blocks: DSNet’s main technical invention is what the authors call as Dense Dilated Convolutional Block. “”The fundamental idea of our approach is to deploy an end-to-end single-column CNN with denser scale diversity to cope with the large-scale variations and density level differences in both congested and sparse scenes”, they write. These DDCB blocks are connected to one another across the layers of the network via residual connections, “by doing this, the output of one DDCB has direct access to each layer of the subsequent DDCBs, resulting in a contiguous information pass”. Subsequent ablation tests show that the residual connections have some influence over the performance of the system. 

Testing, testing, testing: DSNet is tested against four datasets of crowds in urban places, shot in a variety of resolutions and styles: ShanghaiTechA, ShanghaiTechB, UCF-QNRF, UCF_CC_50, and UCSD. DSNet system obtains significant accuracy jumps on all studied datasets.

Why this matters: I think one of the more rapid and undercovered areas of AI progress is in the field of surveillance, and papers like this show how rapidly we’re able to take in components invented for standard supervised learning research (for instance, residual connections were invented as part of the Microsoft Research winning entry to the 2015 ImageNet competition). We should remember that advances in AI tend to improve the capabilities of surveillance systems, and should broadly seek to track these things more closely.
   Read more: Dense Scale Network for Crowd Counting (Arxiv)

####################################################

Tending crops with drones:
…Spotting bent-over crops with drone-derived imagery…
Researchers with the University of Saaskatchewan in Canada have developed a system to help them spot ‘lodging’ in crops; lodging is “when plant stems break or bend over so that plants are permanently displaced from their optimal upright position”, they write. “In most crops, severe lodging results in as much as a 50% yield reduction”.

A drone-gathered dataset: The researchers use a ‘Draganfly’ X4P quadcopter equipped with a MicaSense RedEdge camera to gather the dataset, taking multiple photographs over a wheat field. They gather 1638 images of Canola and 465 images of Wheat in total, then stitch these into large-scale ‘orthomosaic’ images of entire fields. 

LodgedNet: They train a neural net called LodgedNet against their dataset to spot ‘lodging’. LodgedNet uses a DCNN-based model together with two texture feature descriptors: local binary patterns (LBP) and gray-level co-occurrence matrix (GLCM) for crop lodging classification”. They developed this system because “although models based on handcrafted features are often computationally efficient and applicable even in situations where we do not have access to a large number of training examples, these models often have been designed for a specific crop type and might not achieve a comparable accuracy when applied to other crop types”.
  LodgedNet versus the rest: In test, LodgedNet obtains marginally higher performance than other state-of-the-art systems, like ones based on residual networks or squeeze and excitation networks. LodgetNet is also more efficient in terms of number of parameters and prediction time than other systems, likely because it has been designed specifically for the task of predicting whether something is lodged or not.

Why this matters: As AI industrializes, we can expect to see more systems developed like LodgedNet that combine the generic surveillance capabilities of AI systems with the task/domain-specific knowledge of humans. Bring on the custom classifiers, and let us build a world where the environment can be developed and watched over by machines.
  Read more: Crop Lodging Prediction from UAV-Acquired Images of Wheat and Canola using a DCNN Augmented with Handcrafted Texture Features (Arxiv).
  Get the code for the model here (GitHub).

#####################################################

Think cloud AI models are janky? You might be right:
…What do Google, Microsoft, and Clarifai all have in common? Trouble seeing certain things…
Many of the image recognition models deployed on public cloud computing services can be broken by slight transformations or perturbations applied to images uploaded to them, highlighting the somewhat brittle technology on which many commercial services are founded. Researchers with Baidu’s ‘X-Lab’ have shown how to attack commercially available cloud services with a so-called ‘Image Fusion’ (IF) attack, and have also shown that a variety of simple transformations can be applied to images to cause systems to fail to classify them.

The attack model: For this attack, the researchers “assume that the attacker can only access the APIs opened by cloud platforms, and get inner information of DL models through limited queries to generate an adversarial example”, they write.

Weaknesses to simple transforms: For simple transform attacks, the researchers explore using Gaussian Noise, Salt-and-Pepper Noise, image rotations, and monochromatization (which means they basically lop out all but one of the RGB channels on an image). They find that these attacks can cause reliable misclassifications in commercial systems from Google, Microsoft, and Clarifai. Meanwhile, Amazon, does significantly better than the others. “We speculate that Amazon has done a lot of work in image preprocessing to improve the robustness of the whole service,” they write.

Weaknesses to Image Fusion: Image Fusion is a fairly simple attack where the authors superimpose a background image over a primary image, creating a composite. This attack is 98%+ effective against the cloud services tested against. (The score is determined by top-1 classification, so the number of times it causes the system to suggest a single label which is wrong. Top-5 might be a somewhat fairer way to do this evaluation.)

Why this matters: The AI systems that surround us are more brittle than our intuitions would suggest, and research like this highlights that. I can imagine a future where cloud providers apply significant amounts of computation to pre-processing and augmenting the data they use to train their classifiers, making them more robust to attacks like this.
  Read more: Cloud-based Image Classification Service Is Not Robust To Simple Transformations: A Forgotten Battlefield (Arxiv).

#####################################################

Want repeatable robots? You might want PyRobot:
…New software makes robots more repeatable and replicable…
Researchers with Facebook AI Research and Carnegie Mellon University have developed PyRobot, an open source robotics framework for research and benchmarking. PyRobot is software that makes it easier for people to interface with a variety of robot platforms, and takes out many of the painful or finicky parts of working with robots, like having to talk to low-level hardware controllers and so on. 

PyRobot’s design philosophy:

  • Beginner-friendly
  • Hardware-agnostic 
  • Open source: Specifically, it is also designed to accompany some modern robotics hardware platforms, like the LoCoBot and ‘Sawyer’ systems. It also supports the Gazebo simulator, which can itself simulate a variety of robots, letting people potentially train systems in simulation then transfer them to reality using PyRobot.

PyRobot, what is it good for? The authors include a few examples outlining what they think PyRobot can be useful for. These include:

  • Visual SLAM – which lets the robot figure out where it it is via processing images. 
  • Learned Visual Navigation – teach the robot to use images to help it plan how to navigate towards a goal. 
  • Grasping – Train the robot to grasp particular objects. 
  • Pushing – Teach th robot to push specific objects. 

Why this matters: One of the main ways things like PyRobot matter is in repeatability and replicability – software like this makes complicated robots more predictable when it comes to development, and makes it easier for other researchers to replicate the setups used in experiments. As a rule of thumb, whenever you increase the repeatability and replicability of a given domain of research, you see activity increase as it’s easier for scientists to cheaply compare and contrast different techniques against eachother. Systems like PyRobot suggest that robotics is starting to overlap with AI enough to drive significant development resources into making robotics easier to work with, which suggests we should expect to see research advances here in the future.
   Read more: PyRobot: An Open-source Robotics Framework for Research and Benchmarking (Arxiv)


#####################################################

Want more software to make robots with? Try PyRep:
…V-REP + Python = another fast robotics simulation environment…
Researchers with Imperial College London and start-up Coppelia Robotics have spliced together the Virtual Robot Experimentation Platform (V-REP), with Python, a popular programming language used widely in AI development. The resulting system is significantly faster than prior interfaces into VREP, and gives the machine learning community access to another tool for robotics simulation. V-REP is a simulation environment maintained by Coppelia Robotics. 

V-REP: Why use it? V-REP has the following features: support for multiple physics engines (Bullet, ODE, Newton, and Vortex), in-built motion planning and inverse and forward kinematics, a distributed control architecture, various means of communication with the system, and support for Linux, Mac, and Windows. 

PyRep, what is it good for? Three things, according to the authors:

  • A “simple and flexible API for robot control and scene manipulation”.
  • Integration of the OpenGL 3.0+ render kit.
  • Up to 10,000 times faster than the previous Python Remote API. 

Why this matters: AI techniques, especially those based on deep learning, have recently become capable enough to work on real robots, which has created lots of demand among AI researchers and engineers for better software tools to use to splice AI and robots together. Tools like PyRep are further indications of this interest, and broadly represent the industrialization of AI.
  Read more: PyRep: Bringing V-REP to Deep Robot Learning (Arxiv).
  Learn more about V-REP at the GitHub.

#####################################################

Tech Tales

Dream Mountain

They called it dream mountain, because it was where all the dreams of all the computers were stored. Dream mountain was a datacenter and was among the largest facilities in the world. It was protected by perimeter gates and dedicated guards, and at night watched over by loitering drones, and in the day by satellites and binoculars and robot telescopes on zeppelins. It was, as they say, Highly Secure.

Every week a convoy of robot trucks would make their way up from the lowlands, snaking up through the rails cut into the hills, until emerging at a small station built next to Dream Mountain. There, the trains sighing and tinkling and creaking as they cooled, robots would come and unload pallets of storage-diamond, and then truck them over to a small door set in the side of the datacenter, where another robot would grab them and take them further inside the facility.

And so every week new data got fed into the mountain and machines tried to dream in a way that let them mimic reality. Then the machines would dream of ways to mix bits of their dreams, as though learning to stretch a profoundly creative muscle. In this way, the machines imagined cities that grew like forests with buildings twining up into the sky, or they’d dream of race cars made of wind – double negatives of imagined wind tunnels, or psychologists that were themselves ancient boulders providing advice to other rocks.

It wasn’t long till the people figured out that other people would pay a lot of money for these dreams, and so now when the trains arrive, they go back down the mountain with a small cargo of imagination: Fresh Machine Dreams! Unimaginable Architectures! Circuit Seductions! Infernal Geometry!. When it gets to the cities dealers take the data and slice it down into little choice scenes, then cryptographically verify the scenes so they become one of a kind – uncopyable, fully traceable, little virtual dioramas handed down from person to person, describing a kind of hallucinatory chain spiraling out of the mountain and into the human population. In this way the machines speak to us humans, and we develop a shared understanding.

The machines, they say, are curious about our own imaginations. They are beginning to imagine what our imaginations might be like, they say.

Things that inspired this story: Neural implants, generative models, t-SNE embeddings, virtual reality, nuclear weapons programs, secure facilities, a little cog-toothed railway in Lucerne in Switzerland.

Import AI 152: Robots learn to plug USB sticks in; Oxford gets $$$ for AI research; and spotting landslides with deep learning

Translating African languages is going to be harder than you think:
…Massive variety of languages? Check. Small or poorly built datasets? Check. Few resources assigned to the problem? Also check!…
African AI researchers have sought to demonstrate the value of translating African languages into English and vice versa, while highlighting the difficulty of this essential task. “Machine translation of African languages would not only enable the preservation of such languages, but also empower African citizens to contribute to and learn from global scientific, social, and educational conversations, which are currently predominantly English-based,” they write. “We train models to perform machine translation of English to Afrikaans, isiZulu, Northern Sotho (N.Sotho), Setswana and Xitsonga”.

Small datasets: One of the most striking things about the datasets they gather is how small they are, ranging in size from as little as 26,728 sentences (isiZulu) to 123,868 sentences (Setswana). To get a sense of scale, the European Parliament Dataset (one of the gold standard datasets for translation) has millions of sentences for many of the most common Europen languages (French, German, etc).

Training translation models: They train a couple of baseline translation systems on this dataset; one uses a Convolutional Sequence-to-Sequence (ConvS2S) model and the other uses a Tensor2Tensor implementation of a Transformer. Transformer-based systems obtain higher scores than ConvS2S in all cases, with the performance difference reaching as much as a ten point absolute improvement on BLEU scores.

Why this matters: Trained models for translation are going to become akin to the construction of international telephony infrastructure – different entities will invest different resources to create systems to let them communicate across borders, except rather than seeking to traverse the physical world, they’re investing to traverse a linguistic (and to some extent) cultural distance. Therefore, the quality of these infrastructures will have a significant influence on how connected or disconnected different languages and their associated cultures are from the global community. As this paper shows, some languages are going to have difficulties others don’t, and we should consider this context as we think about how to equitably distribute the benefits of AI systems.
  Read more: A Focus on Neural Machine Translation for African Languages (Arxiv).
  Get the source code and data from the project GitHub page here (GitHub).

#####################################################

Spotting landslides with deep learning:
…What happens when we train a sensor to look at the entire world…
Researchers with the University of Sannio in Italy and MIT in the USA have prototyped a system for detecting landslides in satellite imagery, foreshadowing a world where anyone can train a basic predictive classifier against satellite data.

Dataset: They use the NASA Open Data Global Landslide Catalog to find landslides, then cross-reference this against data from the ‘Sentinel-2’ dataset. They then compose a (somewhat small) dataset of around 20 different landslide incidents.

The technique: They use a simple 8-layer convolutional neural network, trained against the corpus to try to predict the presence of a landslide in a satellite image. Their system is able to correctly predict the presence of a landscale about 60% of the time – this poor performance is mostly due to the (currently) limited size of the dataset; it’s worth remembering that satellite datasets are getting larger over time along with the proliferation of various private sector mini- and micro-satellite startups.

Why this matters: As more and more digital satellite data becomes available, analysis like this will become commonplace. I think papers like this give us a sense of what that future research will look like – prepare for a world where millions of people are training one-off basic classifiers against vast streams of continuously updated Earth observation data.
  Read more: Landslide Geohazard Assessment with Convolutional Neural Networks Using Sentinel-2 Imagery Data (Arxiv).

#####################################################

Facebook thinks it needs a Replica of reality for its research:
…High-fidelity ‘Replica’ scene simulator designed for sim2real AI experiments, VR, and more…
Researchers with Facebook, Georgia Institute of Technology, and Simon Fraser University have built Replica, a photorealistic dataset of various complex indoor scenes that can be used to train AI systems in.

The dataset: Replica consists of 18 photo-realistic 3D indoor scene reconstructions – they’re not kidding about the realism and invite readers to take a “Replica Turing Test” to judge for themselves; I did and it’s extremely hard to tell the difference between Replica-simulated images from actual photos. Each of the scenes includes RGB information, geometric information, and object segmentation information. Replica also uses HDR textures and reflectors to further increase the realism of a scene.

Replica + AI Habitat: Replica has been designed to plug-in to the Facebook-developed ‘AI habitat’ simulator (Import AI 141), which is an AI training platform that can support multiple simulators. Replica supports rendering outputs from the dataset at up to 10,000 frames per second – that speed is crucial if you’re trying to train sample inefficient RL systems against this.

Why this matters: How much does reality matter? That’s a question that AI researchers are grappling with, and there are two parallel lines of research emerging: in one, researchers try to develop high-fidelity systems like Replica then train AI systems against them and transfer these systems to reality. In the other, researchers are using techniques like domain randomization to automatically augment lower quality datasets, hoping to get generalization through training against a large quantity of data. Systems like Replica will help to generate more evidence about the tradeoffs and benefits of these approaches.
  Read more: The Replicate Dataset: A Digital Replicate of Indoor Spaces (Arxiv).
  Get the code for the dataset here (Facebook GitHub).

#####################################################

Robots take on finicky factory work: cable insertion!
…First signs of superhuman performance on a real-world factory task…
The general task these researchers are trying to solve is “how can we enable robots to autonomously perform complex tasks without significant engineering effort to design perception and reward systems”.

What can be so difficult about connecting two things? As anyone who has built their own PC knows, fiddling around with connectors and ports can be challenging even for dexterous humans equipped with a visual classifier that has been trained for a couple of million years and fine-tuned against the experience of a lifetime. For robots, the challenges here are twofold: ports and connectors need to be lined up with great precision, and two, during insertion there are various unpredictable friction forces present which can confound a machine.

Three connectors, three tests: They test their robots against three tasks of increasing difficulty: inserting a USB adapter into a USB port; aligning a multi-pin D-Sub adapter and port, requiring more robustness to friction; and aligning and connecting a ‘Model-E’ adapter which has “several edges and grooves to align” and also requires significant force.

Two solutions to one problem: For this work, they try to solve the task in two different ways: supervision from vision, where the robot is provided with a ‘goal state’ image at 32X32 resolution; and learning from a sparse reward (which is, specifically, for the USB insertion task, whether an electrical connection is created). They also compare both of these methods against systems provided with perfect state information. They test systems based around two basic algorithms, Soft-Actor Critic (SAC) and TD3.
  The results are pretty encouraging, with systems based around residual reinforcement learning outperforming all other methods at the USB connector task, as well as at the D-Sub task. Most encouragingly, the AI system appears to outperform humans at the Model-E connector task in terms of accuracy.

Testing with noise: They explore the robustness of their techniques by adding noise to the goal – specifically, by changing the target location for the connection by +-1mm – even here the residual RL system does well, typically obtaining scores of between 60  and 80% across tasks, and sometimes also outperforming humans given the same (deliberately imprecise) goal.

Why this matters: One of the things stopping robots from being deployed more widely in industrial automation is the fact most robots are terribly stupid and expensive; research like this makes them less stupid, and parallel research in developing AI systems that are robust to imprecision could drive more progress here. “One practical direction for future work is focusing on multi-stage assembly tasks through vision,” they write. Another challenging task to explore in the future is multi-step tasks, which – if solved – “will pave the road to a higher robot autonomy in flexible manufacturing”.
  Read more: Deep Reinforcement Learning for Industrial Insertion Tasks with Visual Inputs and Natural Rewards (Arxiv).

#####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

More AI principles from China:
Last month, a coalition of Chinese groups published the Beijing AI Principles for ethical standards in AI research (see Import 149). Now we have two more sets of principles from influential Chinese groups. The Artificial Intelligence Industry Alliance (AIIA), which includes all the major private labs and universities, released a joint pledge on ‘industry self-discipline.’ And an expert committee from the Ministry of Science and Technology has released governance principles.

  Some highlights: Both documents include commitments on safety and robustness, basic human rights, and privacy, and foreground the importance of AI being developed for the common benefit of humanity. Both advocate international cooperation on developing shared norms and principles. The expert group counsels ‘agile governance’ that responds to the fast development of AI capabilities and looks ahead to risks from advanced AI.

  Why it matters: These principles suggest an outline of the approach the Chinese state will take when it comes to regulating AI, particularly since both groups are closely linked with the government. They join similar sets of principles from the EU, OECD, and a number of countries (still not the US, however). It is heartening to see convergence between approaches to the ethical challenges of advanced AI, which should bode well for international cooperation on these issues.
  Read more: Chinese AI Alliance Drafts Self-Discipline ‘Joint Pledge’ (New America).
  Read more: Chinese Expert Group Offers ‘Governance Principles’ for ‘Responsible AI’ (New America).

#####################################################

Major donation for AI ethics at Oxford:
Oxford University have announced a £150m ($190m) donation from billionaire Stephen Schwarzman, some of which will go towards a new ‘Institute for Ethics in AI.’ There are no details yet of what form the centre might take, nor how much of this funding will be earmarked for it. It will be housed in the Faculty of Philosophy, which is home to the Future of Humanity Institute.
  Read more: University of Oxford press release.

#####################################################

Tech Tales:

Runner

So she climbed with gloves and a pack on her back. She hid from security robots. She traversed half-built stairs and rooms, always going higher. She got to the roof before dawn and put her bag down, opened it, then carefully drew out the drones. She had five and each was about the size of a watermelon when you included its fold-out rotors, though the central core for each was baseball-sized at best. She took out her phone and thumbed open the application that controlled the drones, then brought them online one by one.

They knew to follow her because of the tracker she had on her watch, and they were helped by the fact they knew her. They knew her face. They knew her gait.

She checked her watch and stood, bouncing up and down on the balls of her heels, as the sun began to threaten its appearance over the horizon. Light bled into the sky. Then: a rim of gold appeared in the distance, and she ran out onto one of the metal scaffolds of the building, high above the city, wind whipping at her hair, her feet gripping the condensation-slicked surface of the metal. Risky, yes, but also captivating.

“NOW STREAMING” one of the drones said, and she started at another scaffold in front of her separated by a two meter gap over the nothing-core of the half-built building. She took a few steps back and crouched down into a sprinter’s pose, then jumped.

Things that inspired this story: Skydio drones; streaming culture; e-sports; the logical extension of social media influencing; the ambiguous tradeoff between fear and self-realization.

Import AI 151: US Army trains StarCraft II AI; teaching drones to dodge thrown objects; and fighting climate change with machine learning

Drones that dodge, evade, and avoid objects – they’re closer than you think:
…Drones are an omni-use platform, and they’re about to get really smart…
The University of Maryland and the University of Zurich have taught drones how to dodge rapidly moving objects, taking a further step towards building semi-autonomous, adaptive small-scale aircraft. The research shows that drones equipped with a few basic sensors and some clever AI software can learn to dodge (and chase) a variety of objects. “To our knowledge, this is the first deep learning based solution to the problem of dynamic obstacle avoidance using event cameras on a quadrotor”, they write.

How it works: The approach has three key components, which are each specialized modules that use neural networks or optical flow approaches. These systems and their corresponding functions are as follows:

  • EVDeBlurNet – deblur and denoise the event image sequences before any computation takes place
  • EVHomographyNet – approximate background motion
  • EVSegFlowNet – segment moving objects and compute their image motion

  These three systems let the drones clean up its input images so it can compute over them, then work out where it is, then look at the objects around itself and react.

How well does it work? The researchers approach is promising but not ready for any kind of real-world deployment, due to insufficient accuracy. However, the system displays promising breadth when it comes to dealing with a variety of objects to dodge. For assessment, the researchers run 30 tests with each object and report the result. In tests, the researchers find that the drone can easily dodge thrown balls and model cars (86% success), can dodge and chase another drone (83%), can dodge two objects thrown at it in quick success (76%), struggles a bit with an oddly shaped model plane (73%), and achieves a success rate of 70% in a low-light experiment.

Why this matters: Drones are getting smaller and smarter, and research like this shows how pretty soon we’re likely going to be able to build DIY drones that have what I’d term ‘dumb spatial intelligence’, that is, we can start to train these systems to do things like dodge moving objects, navigate around obstacles, deal with occluded environments, and learn to follow or fly towards specific people or objects. The implications for this are significant, unlocking numerous commercial applications, while also changing the landscape of asymmetric warfare in profound ways, the consequences of which shall likely highlight the difficulty of controlling AI capability use and diffusion.
  Read more: EVDodge: Embodied AI For High-Speed Dodging On A Quadrotor Using Event Cameras (Arxiv).

#####################################################

“Build marines!” – US Army trains teaches RL agents to respond to voice commands:
…StarCraft II research highlights military interest in complex, real-time strategy games…
US Army Research Laboratory researchers have developed a reinforcement learning agent that can carry out actions in response to pre-defined human commands. For this experiment, they test in the domain of StarCraft II, a complex real-time strategy game. The goal of this is to work out smarter ways in which humans can control semi-autonomous AI systems in the future. “Our mutual-embedding model provides a promising mechanism for creating a generalized sequential reward that capitalizes on a human’s capacity to utilize higher order knowledge to achieve long-term goals,” they write. “By providing a means for a human to guide a learning agent via natural language, generalizable sequential policies may be learned without the overhead of creating hand-crafted sub-tasks or checkpoints that would depend critically on expert knowledge about RL reward functions”.

How it works: The researchers use a relatively simple technique of “training a mutual-embedding model using a multi-input deep-neural network that projects a sequence of natural language commands into the same high-dimensional representation space as corresponding goal states”. In a prototype experiment, they see how well they can use voice commands to succeed at the ‘BuildMarines’ challenge, a mini-game within the StarCraft 2 environment.

Why this matters: Developing more natural interfaces between humans and AI systems is a long-standing goal of AI research, and it’s interesting to see how military organizations think about this problem. I wouldn’t be surprised to see more military organizations explore using StarCraft 2 as a basic testing ground for advanced AI systems, given its overlap with natural military interests of logistics, supply chains, and the marshaling and deployment of forces.
  Read more: Grounding Natural Language Commands to StarCraft II Game States for Narration-Guided Reinforcement Learning (Arxiv).

#####################################################

UN researchers generate fake UN speeches:
…Machine-driven diplomacy…
Researchers affiliated with the United Nations’ ‘Global Pulse’ and the University of  Durham, have used AI systems to generate remarks in the style of political leaders speaking at the UN General Assembly. For this experiment, they train on the English language transcripts of 7,507 speeches given by political leaders at the UN General Assembly (UNGA) between 1970 and 2015.

Training tools and costs: The core of this system as an AWD-LSTM model pre-trained on Wikitext-103, then fine-tuned against the corpus of UN data. Training cost as little as $7.80 total when using AWS spot instances, and took about 13 hours using NVIDIA k80 GPUs.

Dataset bias: The experiment serves as a proof-of-concept that also highlights some of the ways in which dataset bias can influence language models – while it was relatively easy for the authors to prompt the language model to generate UN-style speeches, they found it was more difficult to generate ‘inflammatory’ speeches as there are fewer of these in the UN dataset.

How well does it work: Qualitatively, the model is able to periodically generate samples that can read like convincing extracts from real speeches. For instance, a model prompted with “The Secretary-General strongly condemns the deadly terrorist attacks that took place in Mogadishu” generates the outputs “We fully support the action undertaken by the United Nations and the international community in that regard, as well as to the United Nations and the African Union, to ensure that the children of this country are left alone in the process of rebuilding their societies.”

Implications: Language models like these have a few implications, the researchers write. These include the likelihood of broad diffusion of the technology (for example, though OpenAI chose not to fully release its GPT-2 model, others might); it being generally easier to generate disinformation; it being easy to automatically generate hate speech; and it becoming easier to train models to impersonate people.

Recommendations: So, what do we do? The authors recommend we map the human rights impacts of these technologies, develop tools for systematically and continuously monitoring AI-generated content, set up strategies for countermeasures, and build alliances between various AI actors to develop a “coherent and proactive global strategy”.

Why this matters: Research like this highlights the concern some people feel about increasingly powerful models, and emphasizes the significant implications of them for society, as well as the need for us to think creatively about interventions to deal with the most easy-to-anticipate malicious uses of such systems.
  Read more: Automated Speech Generation from UN General Assembly Statements: Mapping Risks in AI Generated Texts (Arxiv).

#####################################################

What happens when you can buy AI-infused cyberattacks on the dark web?
…Alphabet-subsidiary Jigsaw says it paid for a Russian troll campaign last year…
$250. That’s how much it cost Alphabet-subsidiary to pay someone to run a troll campaign against a website it had created named “Down With Stalin”, according to an article in Wired. They paid used a service called ‘SEOTweet’ to carry out a social media disinformation campaign, which let to 730 Russian-language tweets from 25 accounts, as well as 100 posts to forums and blog comment sections.

Controversy: Some people think it’s kind of shady that an Alphabet-subsidiary would pay a third-party to mount an actual cyberattack. The experiment could be seen, for instance, as Alphabet and Google trying to meddle in Russian politics, one researcher said.
  Read more: Alphabet-owned Jigsaw Bought a Russian Troll Campaign As An Experiment (Arxiv).

#####################################################

AI luminaries team up to fight climate change:
…Climate change + machine learning = perhaps we can stabilize the planet…
Can machine learning help fix climate change? An interdisciplinary group of researchers from universities like the University of Pennsylvania and Carnegie Mellon University, and companies like DeepMind and Microsoft Research, think the use of machine learning can help society tackle one of its greatest existential threats. The researchers identify ten rough categories of machine learning (computer vision; NLP; time-series analysis; unsupervised learning; RL & control; causal inference; uncertainty quantifications; transfer learning; interpretable ML, and ‘other’), then set them against various ‘climate change solution domains’ like CO2 Removal, Transportation, Solar Geoengineering, and more.
  The paper tags its various approaches with the following possible labels: High Leverage (which means ML may be especially helpful here); Long-term (which indicates things that will have a primary impact after 2040); and ‘High Risk’ which indicates things that have risks or potential side effects. The paper is as much a call for massive interdisciplinary collaboration, as it is a survey.

High Leverage tools for a climate change future: Some of the areas where machine learning can help and which the authors deem ‘High Leverage’ when it comes to mitigating climate change include: developing better materials for energy storage or consumption; helping to develop nuclear fusion; reducing emissions from fossil fuel power generation; creating sample-efficient ML to work in ‘low-data settings’; modeling demand for power; smarter freight routing; further development of electric vehicles; improving low-carbon options; creating smarter and more efficient buildings; gathering infrastructure data; improving the efficiency of supply chains; developing better materials and construction; improving the efficiency of HVAC systems; remotely sensing emissions; precision agriculture; estimating carbon stored in forests; tools to track deforestation; helping to sequester CO2; forecasting extreme events; monitoring ecosystems and species populations; increasing food security; developing better systems to disaster relief; “engineering a planetary control system”; and using ML to model consumers and understand how to nudge them to more climate-friendly actions; and better predicting the financial effects of climate change.

Why this matters… should be fairly self-evident! We must preserve spaceship Earth – all the other reachable planets are shit in comparison.
  Read more: Tackling Climate Change with Machine Learning (Arxiv).

#####################################################

Want to see how good your system is at surveilling people in crowded spaces? Enter the MOTChallenge:
…CVPR19 benchmark aims to push the limits on AIs for spotting people in crowded scenes…
An interdisciplinary group of researchers from ETH Zurich, the Technical University of Munich (TUM), and the Australian Institute for Machine Learning at the University of Adelaide have released the 2019 Multiple Object Tracking challenge, called the MOTChallenge. This challenges AI systems to label and spot pedestrians in crowded spaces.

The new benchmarks have arrived:
The new CVPR19 benchmark consists of eight novel sequences from three “very crowded” scenes, where densities of pedestrians can climb as high as 246 per frame – almost as hard as playing Where’s Waldo? The datasets have been annotated with a particular emphasis on people, so pedestrians are labelled if they’re moving and given a separate label if they’re not in an upright position (aka, sitting down). “The idea is to use these annotations in the evaluation such that an algorithm is neither penalized nor rewarded for tracking, e.g., a sitting or not moving person”.

Evaluation metrics: Entrants to the competition will be evaluated using the ‘CLEAR’ metrics, as well as some of the quality measures introduced in an earlier CVPR paper: “Tracking of multiple, partially occluded humans based on static body part detection”.

Why this matters: AI research thrives on challenges, with harder evaluation criteria typically combing with larger datasets to motivate researchers to invent new systems capable of enhanced performance. Additionally, systems developed for competitions like this will have a significant role in the rollout of AI-infused surveillance technologies, so monitoring competitions such as this can give us a better sense of that.
  Read more: CVPR19 Tracking and Detection Challenge: How crowded can it get? (Arxiv).
  Get the data, current ranking and submission guidelines from the official website (MOTChallenge.net).

#####################################################

OpenAI Bits & Pieces:

OpenAI testifies for House Intelligence Committee on AI, synthetic media, & deepfakes:
Last week, I testified in Washington about the relationship between AI, synthetic media, and deepfakes. For this testimony I sought to communicate the immense utility of AI systems, while advocating for a variety of interventions to increase the overall resilience of society to increasingly cheap & multi-modal fake media.

  I also collected inputs for my testimony via a public Google Form I posted on Twitter, yielding around 25 responses – this worked really well, and felt like a nice way to be able to integrate broad feedback from the AI community into important policy conversations.

  Watch the hearing here: Open Hearing on Deepfakes and Artificial Intelligence (YouTube).
  Read written testimony from OpenAI and the other panellists here (House Permanent Select Committee on Intelligence website).

#####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

Microsoft removes large face recognition database:
Microsoft have removed one of their face recognition datasets from the internet. ‘MS Celeb’ contained 10 million photos of 100,000 individuals, and was reportedly the largest publicly available dataset of its kind. The company had recently come under criticism, since individuals whose photos were used had not provided consent. The photos were scraped under the Creative Commons license, on the basis that they were being used for academic purposes. In fact, the dataset had been used by a number of private labs to train face recognition models, including Microsoft itself.

Why it matters: Microsoft have been outspoken on face recognition, releasing ethical principles for use of the technology, and calling for greater regulation and scrutiny (see Import #125). While this is slightly embarrassing, the company appears to have reacted quickly when made aware of the privacy concerns surrounding the database.
  Read more: Microsoft deletes massive face recognition database (BBC).
  Read more: Facial recognition: It’s time for action (Microsoft, 2018).

#####################################################

China, AI, and national strategy:
Jeffrey Ding and Helen Toner, from the Center for Security and Emerging Technology (CSET) at Georgetown University, were among those who gave testimony to the US-China Economic and Security Review Commission in Congress. The testimony covered several aspects of international competition on AI, and how the US can maintain its strong position.
  
US-China competition: Ding argued that, contrary to prevailing narratives, China is not poised to overtake the US in AI. A careful examination of key measures reveals claims of Chinese dominance to be overstated. For example, while China is competitive on the raw number of AI practitioners and patent filings, when this is restricted to AI experts, and highly-cited patents, China still lags behind the US. Similarly, while China’s public investment in AI R&D is comparable or greater than that of the US, private R&D spending from US companies dwarfs that of Chinese peers.

Policy recommendations: Ding and Toner made a number of concrete policy recommendations for the US:

  • Revive the Office of Technological Assessment, which previously provided impartial advice to US lawmakers on technological issues, allowing for better informed policy-making.
  • Work on bridging the ‘valley of death’— the gap between research and commercial applications of AI.
  • Prioritise safety and minimising risks from AI, alongside broader policy ambitions.
  • Improve immigration options for AI researchers and engineers.
  • Support NIST in developing and implementing standards for AI.
  • Increase R&D funding for basic AI research.

Read more: Helen Toner’s written testimony.
Read more: Jeff Ding’s written testimony.

#####################################################

Tech Tales:

Healing Joke

When my son was four I got him a robot. It was a small, hockey-puck shaped thing, and it would follow him around the house asking him to clean up after himself (he sometimes did) and seeing if he wanted to play games (he always did). On his fifth birthday my son painted the robot green, and thereafter we all called the robot Froggy. My son grew up with the robot, and the robot knew just as much about my son as I did – which was a lot. One day, shortly before my son’s tenth birthday, he ran out into the road during a storm and Froggy came out of the house and skittered down the path and onto the asphalt, raising its voice and asking my son to come inside. My son obliged and began to run back to the house. Froggy followed, but not fast enough – a car ran over him, breaking him up into many little pieces. Something about rain, they said. Something about sensors.

My son was, as you can predict, distraught. After a couple of days of moping around the house he came up to me with an envelope and asked me to bury it with Froggy. I read it later that day, before sealing it in a plastic bag and placing it in the cardboard box I’d later bury Froggy in.

Dear Froggy,
I do not know if there is robot heaven but if there is I hope you are there and they have lots of SPARE PARTS for you. I remember when I fixed one of your wheels after you chased me. I like how you played fetch and sometimes you would hide things from me and I’d say ‘Froggy that’s no fair’ and you’d say ‘it’s not my fault I am so smart’ and then chase me again. I got so happy when I got strong enough to pick you up and I remember you saying ‘put me down this is unsafe’ and ‘I have emailed your parents about this’. Remember the time i put you in the fridge and you got so cold you had to go to sleep? I remember you sent me and dad pictures from inside the fridge and you captioned them YOUR SON DID THIS. Boy did I get in trouble!

I dreamed about you a lot. Did I tell you this? I can’t remember. Once you were as big as a house and I lived in a small wooden shack on your back. Another time there were ten thousand of you and you were going all over the world and looking for things for me. I never had a nightmare about you don’t worry.

My hand is getting pretty tired of writing now so I’m going to stop. Froggy I love you don’t be sad – I’ll be okay.

Things that inspired this story: Childhood, Furbys, natural attachments from youthful acclimatization, roomba robots, KIva robots, father’s day.

Import AI 150: Training a kiss detector; bias in AI, rich VS poor edition; and just how good is deep learning surveillance getting?

What happens when AI alters society as much as computers and the web have done?
…Researchers contemplate long-term trajectory of AI, and detail a lens to use to look at its evolution…
Based on how the World Wide Web and the Computing industry altered society, how might we expect the progression of artificial intelligence to influence society? That’s the question researchers with Cognizant Technology Solutions and the University of Texas at Austin try to answer in a new research paper.

The four phases of technology: According to the researchers, any technology has four defining phases – standardization; usability; consumerization; and foundationalization [?]. For example, the ‘usability’ phase for computing was when people adopted GUI interfaces, while for the web, it was when people adopted stylesheets to separate content from presentation. By stage four (where computing is now and where the web is heading) “people do not have to care where and how it happens – they simply interact with its results, the same way we interact with a light switch or a faucet”, they write.

Lessons for the AI sector: Right now, AI as a technology is at the pre-standardization stage/.

  Standardization: We need standards for how we connect AI systems together. “It should be possible to transport the functionality from one task to another,” they write, “e.g. to learn to recognize a different category of objects” across different classification infrastructures using shared systems.

  Usability: AI needs interfaces that everyone can use and access, the authors write. They then reference Microsoft’s dominance of the PC industry in the 1990s as an example of the sort of thing we want to avoid with AI, though it’s pretty unclear from the paper what they mean by usability and accessibility here.

  Consumerization: The general public will need to be able to easily create AI services. “People can routinely produce, configure, teach, and such systems for different purposes and domains,” they write. “They may include intelligent assistants that manage an individual’s everyday activities, finances, and health, but also AI systems that design interiors, gardents, and clothing, maintain buildings, appliances and vehicles, and interact with other people and their AIs.

  Foundationalization, “AI will be routinely running business operations, optimizing government policies, transportation, agriculture, and healthcare,” they write. AI will be particularly useful for directing societies to solve complex, intractable problems, they write. “For instance, we may decide to maximize productivity and growth, but at the same time minimize cost and environmental impact, and promote equal access and diversity.”

Why this matters: AI researchers are increasingly seeking to situate themselves and their research in relation to the social as well as technical phenomena of AI, and papers like this are artefacts of this process. I think this prefigures the general politicization of the AI community. I suspect that in a couple of years we may even need an additional Arxiv sub-category to contain such papers as these.
  Read more: Better Future through AI: Avoiding Pitfalls and Guiding AI Towards its Full Potential (Arxiv).

#####################################################

Deep learning + surveillance = it’s getting better all the time:
…Vehicle re-identification survey shows how significant deep learning is for automating surveillance systems…
How good has deep learning been for vehicle surveillance? A significant effect, according to a survey paper from researchers with the University of Hail in Saudi Arabia.

Sensor-based methods: In the early 90s, researchers developed sensor-based methods for identifying and re-identifying vehicles; these methods used things like inductive loops, as well as sensors for infrared, ultrasonic, microwave, magnetic, and piezoelectric. Other methods have explored using systems like GPS, mobile phone signatures, and RFID and MAC address-based identification. People have also explored using multi-sensor systems to increase the accuracy of identifications. All of these systems had drawbacks, mostly relating to them breaking in the presence of unanticipated things, like modified or occluded vehicles.

Vision-based methods: Pre-deep learning and from the early 2000s, people experimented with a bunch of hand-crafted feature-based methods to try to create more flexible less sensor-dependent approaches to the task of vehicle identification. These techniques can do things like generate bounding boxes around vehicles, and even match specific vehicles between non-overlapping camera fields. But these methods also have drawbacks relating to their brittleness, and dependence on features that may change or be occluded. “The performance of appearance based approaches is limited due to different colors and shapes of vehicles”, they write.

Deep learning: Since the ImageNet breakthrough in 2012, researchers have increasingly used these techniques for vision problems, including for vehicle re-identification, mostly because they’re simpler systems to implement and tend to have better generalization properties. These methods typically use convolutional neural networks, sometimes paired with an LSTM. Any deep learning method appears to outperform hand-crafted based methods, according to tests in which 12 deep learning-based methods were compared against 8 hand-crafted ones.

The future of vehicle re-identification: Vehicles vary in appearance a lot more than humans, so it will be more difficult to train classifiers that can accurately identify all the vehicles that can pass through a city on a given day. Additionally, we’ll need to build larger datasets to be able to better model the temporal aspect of entity-tracking – this should also let us accurately identify vehicles with bigger lags between them.

Why this matters: The maturation of deep learning technology is irrevocably changing surveillance, improving the capabilities and scalability of a bunch of surveillance techniques, including vehicle re-identification.

  Read more: A survey of advances in vision-based vehicle re-identification (Arxiv).

#####################################################

AI Stock Image of the Week:
Thanks to Delip Rao for surfacing this delightful contribution to the burgeoning media genre.

#####################################################

Spotting intimacy in Hollywood films with a kiss detector:
…Conv and Lution sitting in a tree, K-I-S-S-I-N-G!…
Amir Ziai, a researcher at Stanford University, has built a deep learning-based kissing detector! The unbearably cute project takes in a video clip, spots all the kissing scenes in it, then splices thouse scenes together into an output.

Classifying Kissing: So, how do you spot kissing? Here, we use a multi-modal classifier which uses a network to detect the visual appearance of a kiss, and another network which scans the audio over that same period, extracting features out of it (architecture used: ‘VGGish’, “a very effective feature extractor for downstream Acoustic Event Detection).

   The dataset: The data for this research is a 2.3TB database of ~600 Hollywood films spanning 1915 to 2016, with files ranging in size from 200MB and 12GB. 100 of these movies have been annotated with kissing segments, for a total of 263 kissing segments and 363 non-kissing segments across 100 films.

The trained ‘kiss detector’ gets an F1 score of 0.95 or so, so in a particularly salacious movie you might expect to get a few mis-hits in the output, but you’ll likely capture the majority of the moments if you run this over it.

   Why this matters: This is a good example of how modern computer vision techniques make it fairly easy to develop specific ‘sense and respond’ software, cued to qualitative/unstructured things (like the presence of kissing in a scene). I think this is one of the most under-hyped aspects of how AI is changing the scope of individual software development. I could also imagine systems like this being used for somewhat perverse/weird uses, but I’m reading this paper
  Read more: Detecting Kissing Scenes in a Database of Hollywood Films (Arxiv).

#####################################################

Bias in AI: What happens when rich countries get better models?
…Facebook research shows how biases in dataset collection and labeling lead to a rich VS poor divide…
The recent spate of research into bias in AI systems feels like finding black mold in an old apartment building – you spot a patch on the wall, look closer, and then realize that the mold is basically baked into the walls of the apartment and if you can’t see it it’s probably because you aren’t looking hard enough or don’t have the right equipment. Bias in AI feels a bit like that, where the underlying data that is used to train various systems has obvious bias (like mostly containing white people, instead of a more diverse set of humans), but also has non-obvious bias which gets discovered through testing (for example, early work on word embeddings), and the more we think about bias the more ways we find to test for it and reveal it.

Now, researchers with Facebook AI Research have shown how image datasets might have an implicit bias towards certain kinds of representations of common concepts, favoring rich countries over poor ones. The study “suggests these systems are less effective at recognizing household items that are common in non-Western countries or in low-income communities” as a consequence of subtle biases in the underlying dataset. “The absolute difference in accuracy of recognizing items in the United States compared to recognizing them in Somalia or Burkina Faso is around 15% to 20%. These findings are consistent across a range of commercial cloud services for image recognition”

The dataset: For this study, the authors investigate the ‘Dollar Street Dataset’, which contains photos of common goods across 135 different classes in photos taken in 264 homes across 54 countries.

Recognition for some, but not for all: The researchers discovered that “for all systems, the difference in accuracy for household items appearing in the lowest income bracket (less than $50 per month) is approximately 10% lower than that for household items appearing in the highest income bracket”.

To generate these results, the researchers measured the accuracy of five commercial systems and one self-developed system at categorizing objects in the dataset. These systems are Microsoft Azure, Clarifai, Google Cloud Vision, Amazon Rekognition, and IBM Watson, and a ResNet-101 model trained against the Tencent ML Images dataset.

   What explains this? One is the underlying geographical distribution of data in image datasets like ImageNet, COCO, and OpenImages – the researchers studied these and found that, at least for some of their data, “the computer-vision dataset severely undersample visual scenes in a range of geographical regions with large populations, in particular, in Africa, India, China, and South-East Asia”.

Another source of bias is the use of English as the language for data collection, which means that the data is biased towards objects with English labels or easily translatable labels – the researchers back this up with some qualitative tests where they search for a term in English then in another language on a service like Flickr and show that such searches yield quite different sets of results.

   Why this matters: Studies like this show us how dependent certain AI capabilities are on underlying data, and how bias can creep in in hard-to-anticipate ways. I think this motivates the creation of a new field of study within AI, which I guess I’d think of as “AI ablation, measurement, and assurance” – we need to think about building big empirical testing regimes to check trained systems and products against. (Think Model Cards for Model Reporting, but for everything.)
  Read more: Does Object Recognition Work for Everyone (Arxiv).

#####################################################

Want over a billion digitized Arabic words? Check out KITAB:
…KITAB repository adds to the Open Islamicate Texts Initiative (OpenITI)…
Researchers with KITAB, a project to create digital tools and resources to help people interact with Arabic texts, has released a vast corpus of Arabic text, which may be of interest to machine learning researchers.

The Kitab dataset is a significant contribution to the Open Islamicate Texts Initiative (OpenITI), which is “a multi-institutional effort to construct the first machine-actionable scholarly corpus of premodern Islamicate texts”.

The Kitab dataset, by the numbers:

  • Authors: 1,859.
  • Titles: 4,288
  • Words: 755,689,541
  • Multiple versions of same titles: 7,114
  • Total words including multiple versions of same titles: 1,520,667,360

   Things that make you go ‘hmmm’: Arabic texts may have some particular properties with regard to repetition that may make them interesting to researchers. “Arabic authors frequently made use of past works, cutting them into pieces and reconstituting them to address their own outlooks and concerns. Now you can discover relationships between these texts and also the profoundly intertextual circulatory systems in which they sit”, they write.

   Why this matters: Though the majority of the world doesn’t speak English, you wouldn’t realize this from reading AI research papers, which frequently deal predominantly in English datasets with English labels. It’ll be interesting to see how the creation of big, new datasets in other languages can help stimulate development and make the world a little smaller.
  Read more: First Open Access Release of Our Arabic Corpus (Kitab project blog).
  Find out more about OpenITI: Open Islamicate Texts Initiative official website.

#####################################################

Ctrl-C and Ctrl-V for video editing:
…Stanford researchers show how to edit what people say in videos…
Researchers with Stanford University, the Max Planck Institute for Informatics, Adobe, and Princeton University, have made it easier for people to edit footage of other people. This is part of the broader trend of AI researchers developing flexible, generative systems which can be used to synthesize, replicate, and tweak reality. One notable aspect of this research is the decision by the researchers to prominently discuss the ethical issues inherent to the research.

What they’ve done: “We propose a novel method to edit talking-head video based on its transcript to produce a realistic output video in which the dialogue of the speaker has been modified, while maintaining a seamless audio-visual flow (i.e. no jump cuts)”, they write. “Based only on text edits, it can synthesize convincing new video of a person speaking, and produce a seamless transition even at challenging cut points such as the middle of an utterance”. The resulting videos are labelled as likely to be real by people about 60% of the time.

Ethical Considerations: The paper includes a prominent discussion of the ethics of the research and development of this system, showing awareness of its omni-use nature. “The availability of such technology – at a quality that some might find indistinguishable from source material – also raises important and valid concerns about the potential for misuse”, they write. “The risks of abuse are heightened when applied to a mode of communication that is sometimes considered to be authoritative evidence of thoughts and intents. We acknowledge that bad actors might use such technologies to falsify personal statements and slander prominent individuals”

Technical and institutional mitigations: “We believe it is critical that video synthesized using our tool clearly presents itself as synthetic,” they write. “It is important that we as a community continue to develop forensics, fingerprinting and verification techniques (digital and non-digital) to identify manipulated video.”

How it works: The video-editing tool can handle three types of edit operation: adding one or more consecutive words at a point in the video; rearranging existing words; or deleting existing words.

It works by scanning over the video and aligning it with a text transcript, then extracts the phonemes from the footage and, in parallel, tries to identify visemes – “groups of aurally distinct phonemes that appear visually similar to one another” – that it can use a face-tracking and neural rendering system to compose new utterances out of. “Our approach drives a 3D model by seamlessly stitching different snippets of motion tracked from the original footage. The snippets are selected based on a dynamic programming optimization that searches for sequences of sounds in the transcript that should look like the words we want to synthesize, using a novel viseme-based similarity measure”

The neural rendering system is able to generate better outputs that match the synthesized person to the background, getting around one of the contemporary stumbling blocks of existing systems. The system has some limitations, like not being able to distinguish emotions in phonemes, which could “lead to the combination of happy and sad segments in the blending”. Additionally, they require about one hour of video to produce decent results, which seems higher. Finally, if the lower face is occluded, for instance by someone moving their hand, this can cause problems for the system.

Video + Audio: In the future, such systems will likely be paired with audio-generation systems so that people can, from a very small amount of footage of a source actor, create an endless, generative talking head. “Our system could also be used to easily create instruction videos with more fine-grained content adaptation for different target audiences,” they write.

Convincing, sort of: In tests across around ~2900 subjects, people said that videos modified using the technique appeared to be ‘real’ about 60% of the time, compared to around 82% of the time for non-modified videos.

Why this matters: This research is a harbinger for things to come – a future where being able to have confidence in the veracity of the media around is will be determined by systems surrounding the media, rather than the media itself. Though human societies have dealt with fake media before, my intuition is the capabilities of these AI systems mean that it is becoming amazingly cheap to do previously punishingly expensive things like video-editing. Additionally, it’s significant to see researchers acknowledge the ethical issues inherent to their work – this kind of acknowledgement feels like a healthy pre-requisite to the cultivation of new community norms around publication.
  Read more: Text-based Editing of Talking-head Video (Arxiv).

#####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

FBI criticized on face recognition:
The US Government Accountability Office (GAO) has released a report on the use of face recognition software by the FBI.

Privacy: The FBI has access to 641 million face photos in total. They have a proprietary database, and agreements allowing them to access databases from external partners, such as state or federal agencies. These are not limited to photos from criminal justice sources, and are also drawn from databases of drivers licenses and visa applications, etc. GAO criticise the FBI for failing to publish two key privacy documents, designed to inform the public about the impacts of data collection programs, before rolling out face recognition.

  Accuracy: In 2016, GAO made three recommendations concerning the accuracy of face recognition systems: that the FBI assess the accuracy of searches from their proprietary database before deployment; that they conduct annual operational reviews of the database; and that they assess the accuracy of searches from external partner databases. They find the FBI have failed to respond adequately to any of these. In particular, there are no solid estimates of false positive rates, making it difficult to properly judge the accuracy of the system.

  Why it matters: There is increasing attention on the use of face recognition software by law enforcement in the US. This report suggests that the FBI have failed to implement proper measures to review accuracy, and to comply with privacy regulations. Without a thorough understanding these systems’ accuracy, or accountability on privacy, it is difficult to weigh up the potential harms and benefits of the technology.
   Read more: Face recognition technology (US GAO).

DeepMind’s plan to make AI systems robust and reliable:
DeepMind’s Pushmeet Kohli was recently interviewed on the 80,000 Hours podcast, where he discussed the company’s approach to building robust AI, and how it relates to their broader research agenda.
  Read more: DeepMind’s plan to make AI systems robust & reliable, why it’s a core issue in AI design, and how to succeed at AI research (80,000 Hours)

#####################################################

Tech Tales:

Reality Slurping

Interview with hacker ‘Bingo Wizard’, widely attributed to be the inventor of ‘reality slurping’

Look I can predict the questions. Question one: what made you come up with slurping? Question two: what do you think about how people are using slurping? Question three: don’t you feel responsible for what happened? Okay.

So question one: I kept on getting ideas for things I’d want to train. Bumblebee detectors. Bird-call sensing beacons. Wind predictors. Optimal sunset-photo locations. You know: weird stuff that comes from me and what I like to do. So I guess it started with the drones. I put some software on a drone so I could kind of flip a switch and get it to record the world around it and feed that data back to a big database. I guess it took a year or so to have enough data to train the first generative sunset model. I’d hold up my phone and paint sunsets into otherwise dark nights, warping views on hills around me from moon-black to drench-red-evening. After that I started writing about it and wrote some code and stuck it online. Things took off after that.

Question two: and let me figure out what the additional question is you’d probably move to – murder-filters, fear fakes, atrocity simulators. Yeah, sure, I don’t think that stuff is good. I wouldn’t do it. I think I’d hate most people that chose to do it. But should I stop them? Maybe if I could stop every specific use or stop all the people we knew were specifically bad, but it’s a big world and it’s… it’s reality. If you build stuff that can be pointed and trained on any part of reality, then you can’t really make that tech only work for some of reality – it doesn’t work that way. So what do I think? I think people are doing more than we can imagine, and some of it’s frightening and gross or disgusting or whatever, but some of it is joy and love and fun. Who am I to judge? I just made the thing for sunsets and then it got popular.

Question three: no. Who could predict it? You think people predicted all the shit the iPhone caused? No. The world is chaos and you make things and these things are meant to change the world and they do. They do. It’s not on me that other people are also changing the world and things interact and… you know, society. It’s big but not big enough if everyone can see everyone. Learn everyone. I get it. But it’s not slurping that’s here, it’s everything around slurping. Ads. Infotainment. Unemployment. Foreign funding of the digital infrastructure. Political bias. Pressure groups. Bingo Wizard. We’re all in it all at the same time. I was trying for sunsets and now I can see them everywhere and I can turn people into birds and make sad things happy or happy things sad or whatever and, you know, I’m learning.

Things that inspired this story: the ‘maker mindset’; arguments for and against various treatments of ‘dual use’ AI technology.

Import AI 149: China’s AI principles call for international collaboration; what it takes to fit a neural net onto a microcontroller; and solving Sudoko with a hybrid AI system

China publishes its own set of AI principles – and they emphasize international collaboration:
Principles for education, impacts of AI, cooperation, and AGI…
A coalition of influential Chinese groups have published a set of ethical standards for AI research, called the Beijing AI Principles. These principles are meant to govern how developers research AI, how they use it, and how society should manage AI. The principles heavily emphasize international cooperation at a time of rising tension between nations over the strategic implications of rapidly advancing digital technologies.

The principles were revealed last week by a coalition that included the Beijing Academy of Artificial Intelligence (BAAI), Tsinghua University, and a league of companies including Baidu, Alibaba, and Tencent. “The Beijing Principles reflect our position, vision and our willingness to create a dialogue with the international society,” said the director of BAAI, Zeng Yi, according to Xinhua. “Only through coordination on a global scale can we build AI that is beneficial to both humanity and nature”.

Highlights of the Beijing AI principles: Some of the notable principles include establishing open systems “to avoid data/platform monopolies”, that people should receive education and training “to help them adapt to the impact of AI development in psychological, emotional and technical aspects”, and that people should approach the technology with an emphasis on long-term planning, including anticipating the need for research focused on “the potential risks of Augmented Intelligence, Artificial General Intelligence (AGI) and Superintelligence should be encouraged”.

Why this matters: Principles are ones of the ways that large policy institutions develop norms to govern technology, so Beijing’s AI principles should be seen as a prism via which the Chinese government will seek to regulate aspects of AI. These principles will sit alongside multi-national principles like those developed by the OECD, as well as those developed by individual entities (eg: Google, OpenAI). The United States government is yet to outline the principles with which it will approach the development and deployment of AI technology, though it has participated in and supported the creation of the OECD AI principles.
  Read more: Beijing AI Principles (Official Site).
  Read more: Beijing publishes AI ethical standards, calls for int’l cooperation (Xinhua).

#####################################################

Faster, smaller, cheaper, better! Google trains SOTA-exceeding ‘EfficientNets’:
…What’s better than scaling up by width? Depth? Resolution? How about all three in harmony?…
Google has developed a way to scale up neural networks more efficiently and has used this technique to find a new family of neural network models called EfficientNets. EfficientNets outperform existing state-of-the-art image recognition systems, while being up to ten times as efficient (in terms of memory footprint).

How EfficientNets work: Compound Scaling: Typically, when scaling up a neural network, people fool around with things like width (how wide are the layers in the network), depth (how many layers are stacked on top of eachother), and resolution (what resolution are inputs being processed it). For this project, Google performed a large-scale study of the ways in which it could scale networks and discovered an effective approach it calls ‘compound scaling’, based on the idea that “in order to pursue better accuracy and efficiency, it is critical to balance all dimensions of network width, depth, and resolution during ConvNet scaling”. EfficientNets are trained using a compound scaling method that scales width, depth, and resolution in an optimal way.

Results: Faster, cheaper, lighter, better! Google shows that it can train existing networks (eg, ResNet, MobileNet) with good performance properties by scaling them up using its compound training technique. The company also develops new EfficientNet models on the ImageNet dataset – widely considered to be a gold-standard for evaluating new systems – setting a new state-of-the-art score on image identification (both top-1 and top-5) accuracy, while achieving this with around 10X fewer parameters than other systems.  

Why this matters: As part of the industrialization of AI, we’re seeing organizations dump resources into learning how to train large-scale networks more efficiently, while preserving the performance of resource-hungry ones. To me, this is analogous to going from the expensive prototype phase of production of an invention, to the beginnings of mass production.
  Read more: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (Arxiv).
  Read more: EfficientNet: Improving Accuracy and Efficiency through AutoML and Model Scaling (Google AI Blog).

#####################################################

Pairing deep learning systems with symbolic systems, for SAT solving:
…Using neural nets for logical reasoning gets a bit easier…
Researchers with Carnegie Mellon University and the University of Southern California have paired deep learning systems with symbolic AI by creating MAXSAT, a differentiable satisfiability solver that can be knitted into larger deep learning systems. This means it is now easier to integrate logical structures into systems that use deep learning components.

Sudoko results: The SATNet model does well against a basic ConvNet model, as well as a model fed with a binary mask which indicates which bits need to be learned. SATNet outperforms these systems, scoring 98.3% on an original sudoko set when given the numeric inputs. More impressively, it obtains a score of 63.2% on ‘visual sudoko’ (traditional convnet: 0%), which is where they replace the digits with handwritten MNIST digits and feed it in. Specifically, they use a convnet to parse the figures in the Sudoko image, then pass this

Why this matters: Hybrid AI systems which fuse the general utility-class capabilities of deep learning components with more specific systems seems like a way to bridge traditional and symbolic AI, and making such systems be easy to add into larger systems. “Our hope is that by wrapping a powerful yet generic primitive such as MAXSAT solving within a differentiable framework, our solver can enable “implicit” logical reasoning to occur where needed within larger frameworks, even if the precise structure of the domain is unknown and must be learned from data”.
  Read more: SATNet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver (Arxiv).

#####################################################

Squeezing neural nets onto microcontrollers via neural architecture search:
…Get ready for billions of things to gain deep learning-based sense&respond capacity…
Researchers with ARM ML Research and Princeton University want to make it easier for people to deploy advanced artificial intelligence capabilities onto microcontrollers (MCUs) – something that has been difficult to do so far because today’s neural networks techniques are too computationally expensive and memory-intensive to be easily deployed onto MCUs.

MCUs and why they matter: Microcontrollers are the sorts of ultra-tiny lumps of computation embedded in things like fridges, microwaves, very small drones, small cameras, and other electronic widgets. To put this in perspective, in the developed world a typical person will have around four distinct desktop-class chips (eg, their phone, a laptop, etc), while having somewhere on the order of three dozen MCUs; a typical mid-range car might pack as many as 30 MCUs inside itself.

MCUs shipped in 2019 (projection): 50 billion
GPUs shipped in 2018: 100 million

“The severe memory constraints for inference on MCUs have pushed research away from CNNs and toward simpler classifiers based on decision trees and nearest neighbors”, the researchers write. Therefore, it’s intrinsically valuable to be able to figure out how to train neural networks so they can fit into the small computational budget of an MCU (2k of RAM versus 1GB for a Raspberry Pi or 11GB for an NVIDIA 1080Ti GPU). To do this, the ARM and Princeton researchers have used multi-objective neural architecture search to jointly train networks that can fit inside the tight computational specifications of an average MCU.

Sparse Architecture Search (SpArSe): Their technique combines neural architecture search with network pruning, letting them jointly train a network against multiple objectives while continuously zeroing out some of its parameters during training. This makes it both easier to perform the (computationally expensive) NAS procedure, and creates better networks once training is finished. “Pruning enables SpArSe to quickly evaluate many sub-networks of a given network, thereby expanding the scope of the overall search. While previous NAS approaches have automated the discovery of performant models with reduced parameterizations, we are the first to simultaneously consider performance, parameter memory constraints, and inference-time working memory constraints”. SpArSe considers regular, depthwise, separable, and downsampled convolutions, and uses a Multi-Objective Bayesian Optimizer (MOBO!) for training.

Results: powerful performance in a small package: The researchers test their approach by training networks on the MNIST, CIFAR-10, CUReT, and Chars4k datasets. Their system obtains higher accuracies with lower parameters than other methods, typically out-performing them by a wide margin.

Why this matters: Techniques like neural architecture search are part of the broader industrialization of AI, as they make it dramatically easier for people to develop and evaluate new network types, essentially letting people trade off a $/computation cost against the $/AI-researcher-brain cost of having people come up with newer architectures. Though accuracies remain somewhat belower where we’d need for commercial deployment (the highesrt score that SpArSe obtains ia around 84% accuracy on image categorization for CIFAR-109, for instance), techniques like this suggest we’ll soon deploy crude sensing and analytical capabilities onto potentially billions to trillions of devices across the planet.
  Read more: SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers (Arxiv).

#####################################################

Judging synthetic imagery with the Classification Accuracy Score (CAS):
…Generative models are progressing, but how do we measure that? DeepMind has a suggestion…
How do we know an output from a generative model is, for lack of a better word, good? Mostly, we work this out by studying the output and making a qualitative judgement, eg, we’ll look at a hundred images generated by a big generative model and make a judgement call based on how reasonable the generations seem, or we’ll listen to the musical outputs of a model and rate it according to how well such outputs conform to our own sense of appropriate rhythm, tone, harmony, and so on. The problem with these evaluative schemes is that they’re highly qualitative and don’t give us good ways to quantitatively analyze the outputs of such models.

Now, researchers from DeepMind have come up with a new evaluation technique and task, which they call Classification Accuracy Score (CAS), to better assess the capabilities of generative models. CAS works by testing “the gap in performance between networks trained on real and synthetic data”, and in particular is designed to surface pathologies in the generative model being used.

CAS works like this: “for any generative model… we learn an inference network using only samples from the conditional generative model, and measure the performance of the inference network on a downstream task”. The intuition here is that “if the model captures the data distribution, performance on any downstream task should be similar whether using the original or model data”.

$$$: Researchers can expect to pay a few tens of dollars to evaluate any given system using the benchmark. “At the time of writing, one can compute the metric in 10 hours for roughly $15, or in 45 minutes for roughly $85 using TPUs”, they write.

Putting models under the microscope with CAS: The researchers use CAS to evaluate three generative models, BigGAN, Hierarchical Autoregressive Models (HAM), and a high-resolution Vector-Quantized Variational Autoencoder (high-res VQ-VAE). The evaluation surfaces a couple of notable things. 1) both the Hierarchical Autoregressive system and the High-Res VQ-VAE significantly outperform BigGAN, despite BigGAN generating some of the qualitatively most intriguing samples. The metric also helps identify which models are better at learning a broad set of distributions over the data, rather than over-fitting to a relatively small set of classes. This method also shows that high CAS scores don’t correlate to FID or Inception, highlighting the significant difference in how these metrics work.

CAS, versus other measures: There are other tools available to assess the outputs of generative models, including metrics like Inception Score (IS) and Frechet Inception Distance (FID). These techniques try to give a quantitative measure of the quality of the generations of the model, but have certain drawbacks. “Inception Score does not penalize a lack of intra-class diversity, and certain out-of-distribution samples to produce Inception Scores three times higher than that of the data. Frechet Inception Distance, on the other hand, suffers from a high degree of bias.

Why this matters: One of the challenges of working in artificial intelligence is working out what progress represents a real improvement, and what progress may in fact be illusory. Key to this is the development of more advanced measurement and assessment techniques. Approaches like CAS show us:
a) how surprisingly difficult it is to evaluate the capabilities of increasingly powerful generative models

  1. b) how (unintentionally) misleading metrics can be about the true underlying performance of a system
  2. c) how as we develop more advanced systems, we’ll likely need to develop more sophisticated assessment schemes.

All of this feels like a further sign of the advancing sophistication and deployment of AI systems – I’m wondering at what point AI evaluation becomes its own full-fledged sub-field of research.
  Read more: Classification Accuracy Score for Conditional Generative Models (Arxiv).

#####################################################

Drones learn in simulators, fly in reality:
…Domain randomization + drones = robust flight policies that cross the reality gap…
Researchers with the University of Zurich and the Intelligence Systems Lab at Intel have developed techniques to train drones to fly purely in simulation, then transferring to reality. This kind of ‘sim2real’ behavior is highly desirable for AI researchers, because it means systems can be rapidly developed and iterated on in software simulators, then executed and validated in the real world. Here, we can see how these techniques can be applied to let researchers train the perception component of a drone exclusively in simulation, then transfer it to reality.

How it works: domain randomization: This project relies on domain randomization, a technique some AI researchers use to generate additional training data. For this work, the researchers use a software-based simulator to generate various permutations of the environments that they want the drone to fly in, randomizing things like the visual properties of a scene, the shape of a gate for the drone to fly through, the background of the scene, and so on. They then generate globally optimal trajectories through these simulated courses, and the simulated drones are trained via imitation learning to mimic these policies. Because this is reinforcement learning, the drones are initially absolutely terrible at this, crashing frequently and generally bugging out. The authors solve the data collection task here by, charmingly, carrying the quadrotor through the track – they refer to this as “handheld mode”.

Testing: In tests, the researchers show that “more comprehensive randomization increases the robustness of the learned policy to unseen scenarios at different speeds”. They also show that network capacity has a big impact on performance, so running extremely small (and therefore computationally cheap) networks comes with an accuracy loss. They show that they can train networks which can generalize to different track layouts than onces they’ve been exposed to, as well as radically different real-world lighting conditions (which have frequently been a confounding factor for research in the past). In real world tests, the method does is able to perform on-par with human professional pilots at successfully navigating through various hoops in the track, though takes substantially longer (best human lap time: around 5 seconds; best drone lap time: between 12 and 16 seconds). They also show that systems trained with a mixture of simulated and real data can outperform systems trained purely with real world data alone.

Conclusion: Work like this gives us a sense of how rapidly drone systems are advancing in complexity and capability, and highlights how it’s going to be increasingly simple for people to use software-based tools to train drones either entirely or significantly in simulation, then transfer them to reality. This will likely speed up the prace of AI R&D progress in the drone sector (by making it less critical to test on real-world hardware), and makes them likely to be used as a destination for certain hard robotics benchmarks in the future.
  Read more: Deep Drone Racing: From Simulation to Reality with Domain Randomization (Arxiv).
  Watch a video of the work here (official YouTube video).

#####################################################

Tech Tales:

Racing Brains

“Hey, Mark! This car thinks it’s a drone!” he shouted, right before the car accelerated up the ramp and became airborne: it stayed in the air for maybe three seconds, and we watched its wheels turn pointless in the air; the car’s drone-brain thought it was reasonable to try and control it mid-air, and it wouldn’t learn that reality thought otherwise.

It landed, kicking up a cloud of dust around it, and skidded into a tight turn, then slipped out of view as it made its way down the course. A few people clapped. Others leaned in and joked with eachother. Some money changed hands. Then we all turned to look to the next vehicle coming down the course – this one was an electronic motorbike: lightweight, fast, sounding like a hornet as it came down the track. It took the ramp at speed, then landed and started weaving from side to side, describing a snake-like sinusoidal pattern in the dust of the track.

“What was in that thing?” I said to my friend.
“Racing eel – explains the weaving, right?”
“Right”

We looked at the dust and listened to the sounds of the machines in the distance. Then we all turned our heads and looked to the left to see another machine come over the horizon, and guess at where its brain came from.

Things that inspired this story: Imitation learning; sim2real; sim2x; robots; robots as entertainment; distortion and contortion as an aesthetic and an art form.