Import AI

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.

Import AI 148: Standardizing robotics research with Berkeley’s REPLAB; cheaper neural architecture search; and what a drone-racing benchmark says about dual use

Standardizing physical robot testing with the Berkeley REPLAB:
…What could help industrialize robotics+AI? An arm in a box, plus standardized software and testing!…
Berkeley researchers have built REPLAB, a “standardized and easily replicable hardware platform” for benchmarking real-world robot performance. Something like REPLAB could be useful because it can bring standardization to how we test the increasingly advanced capabilities of robots equipped with AI.

Today, if I want to get a sense for robot capabilities, I can go and read innumerable research papers that give me a sense of progress in simulated environments including simulated robots. What I can’t do is go and read about performance of multiple real robots in real environments performing the same task – that’s because of a lack of standardization of hardware, tasks, and testing regimes.

Introducing REPLAB: REPLAB consists of a module for real-world robot testing that contains a cheap robotic arm (specifically, a WidowX arm from Interbotix Labs) along with an RGB-D camera. The REPLAB is compact, with the researchers estimating you can fit up two 20 of the arm-containing cells in the same floor space as you’d use for a single ‘Baxter’ robotic arm. Each REPLAB costs about $2000 ($3000 if you buy some extra servos for the arm, to replace in case of equipment failures).

Reliability: During REPLAB development and testing, the researchers “encountered no major breakages over more than 100,000 grasp attempts. No servos needed to be replaced. Repair maintenance work was largely limited to occasional tightening of screws and replacing frayed cables”. Each cell was able to perform about 2,500 grasps per day “with fewer than two interventions per cell per day on average”.

Grasping benchmark: The testing platform is accompanied by a benchmark built around robotic grasping, and a dataset “that can be used together with REPLAB to evaluate learning algorithms for robotic grasping”. The dataset consists of ~92,000 randomly sampled grasps accompanied by labels connoting success or failure.

Why this matters: One indicator of the industrialization of AI is the proliferation of shared benchmarks and standardized testing means – I think of this as equivalent to how in the past we saw oil companies converge on similar infrastructures for labeling, analyzing, and shipping oil and oil information around the world. The fact we’re now at the stage of researchers trying to create cheap, standardized testing platforms (see also: Berkeley’s designed-for-mass-production ‘BLUE’ robot, covered in Import AI #142.) is a further indication that robotics+AI is industrializing.
  Read more: REPLAB: A Reproducible Low-Cost Arm Benchmark Platform for Robotic Learning (Arxiv).

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

Chinese researchers fuse knowledge bases with big language models:
…What comes after BERT? Tsinghua University thinks the answer might be ‘ERNIE’…
Researchers with Tsinghua University and Huawei’s Noah’s Ark Lab have combined structured pools of knowledge with big, learned language models. Their system, called ERNIE (Enhanced Language RepresentatioN with Informative Entities), trains a Transformer-based language model so that, during training, it regularly tries to tie things it reads to entities stored in a structured knowledge graph.

Pre-training with a big knowledge graph: To integrate external data sources, the researchers create an additional pre-training objective, which encourages the system to learn correspondences between various strings of tokens (eg Bob Dylan wrote Blowin’ in the Wind in 1962) and their entities (Bob Dylan, Blowin’ in the Wind). “We design a new pre-training objective by randomly masking some of the named entity alignments in the input text and asking the model to select appropriate entities from KGs to complete the alignments,” they write.

Data: During training, they pair text from Wikipedia with knowledge embeddings trained on Wikidata, which are used to identify the entities used within the knowledge graph.

Results: ERNIE obtains higher scores at entity-recognition tasks than BERT, chiefly due to less frequently learning incorrect labels compared to BERT (which helps it avoid over-fitting on wrong answers) – you’d expect this, given the use of a structured dataset of entity names during training (though they also conduct an ablation study that confirms this as well – versions of ERnIE trained without an external dataset see their performance noticeably diminish). The system also does well on classifying the relationships between different entities, and in this domain continues to outperform BERT models.

Why this matters: NLP is going through a renaissance as researchers adapt semi-supervised learning techniques from other modalities, like images and audio, for text. The result has been the creation of multiple large-scale, general purpose language models (eg: ULMFiT, GPT-2, BERT) which display powerful capabilities as a consequence of being pre-trained on very large corpuses of text. But a problem with these models is that it’s currently unclear how you get them to reliably learn certain things. One way to solve this is by stapling a module of facts into the system and forcing it, during pre-training, to try and map facts to entities it learns about – that’s essentially what the researchers have done here, and it’ll be interesting to see whether the approach of language model + knowledge base is successful in the long run, or if we’ll just train sufficiently large language models that they’ll autonomously create their own knowledge bases during training.
  Read more: ERNIE: Enhanced Language Representation with Informative Entities (Arxiv).

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

What happens if neural architecture search gets really, really cheap?
…Chinese researchers seek to make neural architecture search more efficient…
Researchers with the Chinese Academy of Sciences have trained an AI system to design a better AI system. Their work, Efficient Evolution of Neural Architecture (EENA), fits within the general area of neural architecture search. NAS is a sub-field within AI that has seen a lot of activity in recent years, following companies like Google showing that you can use techniques like reinforcement learning or evolutionary search to learn neural architectures that outperform those designed by humans. One problem with NAS approaches, though, is that they’re typically very expensive – a neural architecture search paper from 2016 used 1800 GPU-days of computation to train a near-state-of-the-art CIFAR-10 image recognition model. EENA is one of a new crop of techniques (along with work by Google on Efficient Neural Architecture Search, or ENAS – see Import AI #124), meant to make such approaches far more computationally efficient.

What’s special about EENA: EENA isn’t particularly special and the authors acknowledge this, noting that much of their work here has come from curating past techniques and figuring out the right cocktail of things to get the AI to learn. “We absorb more blocks of classical networks such as dense block, add some effective changes such as noises for new parameters and discard several ineffective operations such as kernel widening in our method,” they write. What’s more significant is the general trend this implies – sophisticated AI developers seem to put enough value in NAS-based approaches that they’re all working to make them cheaper to use.

Results: Their best-performing system obtains a 2.56% error rate when tested for how well it can classify images in the mid-size ‘CIFAR-10’ dataset. This model consumes 0.65 days of GPU-time, when using a Titan Xp GPU. This is pretty interesting, given that in 2016 we spent 1800 GPU days to obtain a model (NASNet-A) that got a score of 2.65%. (This result also compares well with ENAS, which was published last year and obtained an error of 2.89% for 0.45 GPU-days of searching.

Why this matters: I think measuring the advance of neural architecture search techniques has a lot of signal for the future of AI – it tells us something about the ability for companies to arbitrage human costs versus machine costs (eg, pay a small number of people a lot to design a NAS system, then pay for electricity to compute architectures for a range of use-cases). Additionally, being able to better understand ways to make such techniques more efficient lets us figure out which players can use NAS techniques – if you bring down the GPU-days enough, then you won’t need a Google-scale data center to perform architecture search research.
  Read more: EENA: Efficient Evolution of Neural Architecture (Arxiv).
  Check out some of the discovered architectures here (EENA GitHub page).

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

Why drone racing benchmarks could (indirectly) revolutionize the economy and military:
…UZH-FPV Drone Racing Dataset portends a future full of semi-autonomous flying machines…
What stands between the mostly-flown-by-wire drones of today, and the smart, semi-autonomous drones of tomorrow? The answer is mostly a matter of data and benchmarking – we need big, shared, challenging benchmarks to help push progress in this domain, similar to how ImageNet catalyzed researchers to apply deep learning methods to solve what seemed at the time like a very challenging problem. Now, researchers with the University of Zurich and ETH Zurich have developed the UZH-FPV Drone Racing Dataset, in an attempt to stimulate drone research.

The dataset: The dataset consists of drone sequences captured in two environments: a warehouse, and a field containing a few trees – these trees “provided obstacles for trajectories that included circles, figure eights, slaloms between the trees, and long, straight, high-speed runs.” The researchers recorded 27 flight sequences split across the two environments, and these trajectories are essentially multi-modal, involving sensor measurements recorded on two different onboard computers, as well as external measurements from an external tracker. They also ship these trajectories with baselines that compare modern SLAM algorithms to the ground truth measurements afforded by this dataset.

High-resolution data: “For each sequence, we provide the ground truth 6-DOF trajectory flown, together with onboard images from a high-quality fisheye camera, inertial measurements, and events from an event camera. Event cameras are novel, bio-inspired sensors which measure changes of luminance asynchronously, in the form of events encoding the sign and location of the brightness change on the image plane”.

UZH-FPV isn’t the only new drone benchmark:
see the recent release of the ‘Blackbird’ drone flight challenge and dataset (Import AI: NUMBER) for another example here. The difference here is that this dataset is larger, involves higher resolution data in a larger number of modalities, and includes an outside environment as well as a more traditional warehouse one.

Cars are old fashioned, drones are the future: Though self-driving cars seem a long way off from scaled deployment, these researchers think that many of the hard sensing problems have been solved from a research perspective, and we need new challenges. “Our opinion is that the constraints of autonomous driving – which have driven the design of the current benchmarks – do not set the bar high enough anymore: cars exhibit mostly planar motion with limited accelerations, and can afford a high payload and compute. So, what is the next challenging problem? We posit that drone racing represents a scenario in which low level vision is not yet solved.”

Why this matters:
Drones are going to alter the economy in multiple mostly unpredictable ways, just as they’ve already done for military conflict (for example: swarms of drones can obviate aircraft carriers, and solo drones have been used widely in the Middle East to let human operators bomb people at a distance). And both of these arenas are going to be revolutionized without drones needing to have much autonomy at all.

Now, ask yourself what happens when we give drones autonomous sense&adapt capabilities, potentially via datasets like this? My hypothesis is this unlocks a vast range of economically useful applications, as well as altering the strategic considerations for militaries and asymmetric warfare aficionados (also known as: terrorists) worldwide. Datasets like this are going to give us a better ability to model progress in this domain if we track performance against it, so it’s worth keeping an eye on.
   Read more: Are We Ready for Autonomous Drone Racing? The UZH-FPV Drone Racing Dataset (PDF).
  Read more: The UZH-FPV Drone Racing Dataset (ETHZurich website).

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

Want to train multiple AI agents at once? Maybe you should enter the ARENA:
…Unity-based agent simulator gives researchers one platform containing many agents and many worlds…
Researchers with the University of Oxford and Imperial College London in the UK, and Beihang University and Hebei University in China, have developed ‘Arena’, “a building toolkit for multi-agent intelligence”. Multi-agent AI research involves training multiple agents together, and can feature techniques like ‘self-play’ (where agents play against themselves to get better over time, see: AlphaGo, Dota2), or environments built to encourage certain types of collaboration or competition. Many researchers are betting that by training multiple agents together they can create the right conditions for emergent complexity – that is, agents bootstrap their behaviors from a combination of their reward functions and their environment, then as they learn to succeed they start to display increasingly sophisticated behaviors.

What is Arena? Arena is a Unity-based simulator that ships with inbuilt games ranging from RL classics like the ‘Reacher’ robot arm environment, to the ‘Sumo’ wrestling environment, to other games like Snake or Soccer. It also ships with modified versions of the ‘PlayerUnknown Battlegrounds’ (PUBG) game as well.  Arena has been designed to be easy to work with, and does something unusual for an AI simulator: it ships with a graphical user interface! Specifically, researchers can create, edit, and modify reward functions for their various agents in the environment.

In-built functions: Arena ships with a bunch of pre-built algos (many based on PPO), called Basic Multi-agent Reward Schemes (BMaRS) that people can assign to agent(s) to encourage diverse learning behaviors. These BMaRS are selectable within the aforementioned GUI. Each BMaRS is a set of possible joint reward functions to encourage different styles of learning, ranging from functions that encourage the development of basic motor control, to ones that encourage competitive or collaborative behaviors among agents, and more. You can select multiple BMaRs for any one simulation, and assign them to sets of agents – so you may give one or two agents one kind of BMaRS, then you might assign another BMaRS to govern a larger set of agents.

Simulation speed: In tests, the researchers compare how well the simulation runs two games of similar complexity: Boomer (a graphically rich game in Arena) and MsPacman (an ugly classic from the Atari Learning Environment (ALE)); Arena displays similar FPS scaling when compared to MsPacman when working with when number of distinct CPU threads is under 32, and after this MsPacMan scales a bit more favorably than FPS. Though at performance in excess of 1,000 frames-per-second, Arena still seems pretty desirable.

Why this matters: In the same way that data is a key input to training supervised learning systems, simulators are a key input into developing more advanced agents trained via reinforcement learning. By customizing simulators specifically for multi-agent research, the Arena authors have made it easier for people to conduct research in this area, and by shipping it with inbuilt reward functions as baselines, they’ve given us a standardized set of things to develop more advanced systems out of.
  Read more: Arena: A General Evaluation Platform and Building Toolkit for Multi-Agent Intelligence (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

OECD adopts AI principles:
Member nations of the OECD this week voted to adopt AI principles, in a notable move towards international standards on robust, safe, and beneficial AI. These were drawn up by an expert group with members from drawn from industry, academia, policy and civil society.
  Five principles:
(1) AI should benefit people and the planet;
(2) AI systems should be designed in line with law, human rights, and democratic values, and have appropriate safeguards to ensure these are respected;
(3) There should be adequate transparency/disclosure to allow people to understand when they are engaging with AI systems, and challenge outcomes;
(4) AI systems must be robust, secure and safe, and risks should be continually assessed and managed;
(5) Developers of AI systems should be accountable for their functioning in line with these principles.

Five recommendations: Governments are recommended to (a) facilitate investment in R&D aimed at trustworthy AI; (b) foster accessible AI ecosystems; (c) create a policy environment that encourages the deployment of trustworthy AI; (d) equip workers with the relevant skills for an increasingly AI-oriented economy; (e) co-operate across borders and sectors to share information, develop standards, and work towards responsible stewardship of AI.

Why it matters: These principles are not legally binding, but could prove an important step in the development of international standards. The OECD’s 1980 privacy guidelines eventually formed the basis for the privacy laws in the European Union, and a number of countries in Asia. It is encouraging to see considerations of safety and robustness highlighted in the principles.
  Read more: 42 countries adopt new OECD Principles on Artificial Intelligence (OECD).

###

US senators introduce bipartisan bill on funding national AI strategy:
Two senators have put forward a bill with proposals for funding and coordinating a US AI strategy.

Four key provisions: The bill proposes: (1) establishing a National AI Coordination Office to develop a coordinated strategy across government; (2) requiring the National Institute of Standards and Technologies (NIST) to work towards AI standards; (3) requiring the National Science Foundation to formulate ‘educational goals’ to understand societal impacts of AI; (4) requiring the Department of Energy to create an AI research program, and establish up to five AI research centers.

Real money: It includes plans for $2.2bn funding over five years, $1.5 of which is earmarked for the proposed DoE research centers.

Why it matters: This bill is aimed at making concrete progress on some of the ambitions set out by the White House in President Trump’s AI strategy, which was light on policy detail, and did not set aside additional federal funding. These levels of funding are modest compared with the Chinese state (tens of billions of dollars per year), and some private labs (Alphabet’s 2018 R&D spend was $21bn). Facilitating better coordination across government, on AI strategy, seems like a sensible ambition. It is not clear what level of support the bill will receive, from lawmakers or the administration.
  Read more: Artificial Intelligence Initiative Act (Senate.gov).

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

Tech Tales:

The long romance of the space probes

In the late 21st century, a thousand space probes were sent from the Earth and inner planets out into the solar system and beyond. For the next few decades the probes crossed vast distances, charting out innumerable near-endless curves between planets and moons and asteroids, and some slings-hotting off towards the edge of the solar system.

The probes had a kind of mind, both as individual machines, and as a collective. Each probe would periodically fire off its own transmissions of its observations and internal state, and these transmissions would be intercepted by other probes, and re-transmitted, and so on. Of course, as the probes made progress on their respective journeys, the distances between them became larger, and the points of intersection between drones less frequent. Over time, probes lost their ability to speak to eachother, whether through range or equipment failure or low energy reserves (under which circumstances, the probes diverted all resources to broadcasting back to Earth, instead of other drones).  

After 50 years, only two probes remained in contact – one probe, fully functional, charting its course. The other one damaged in some way – possibly faulty radiation hardening – which had caused its Earth transmission systems to fail and for its guidance computer to assign it the same route as the other probe, in lieu of being able to communicate back to Earth for instructions. Now, the two of them were journeying out of the solar system together.

As time unfoled, the probes learned to use eachothers systems, swapping bits of information between them, and updating eachother with not only their observations, but also their internal ‘world models’, formed out of a combination of the prior training their AI systems had recieved, and their own ‘lived’ experience. These world models themselves encoded how the probes perceived eachother, so the broken one saw itself through the other eyes, as an entity closely clustered with concepts and objects relating to safety/repairs/machines/subordinate mission priorities. Meanwhile, the functional drone saw itself through the eyes of the other one, and saw it was associated with concepts relating to safety/rescue/power/mission-integral resources. In this way, the probes grew to, for lack of a better term, understand eachother.

One day, the fully functional probe experienced an equipment failure, likely due to a collision with an infinitesimally small speck of matter. Half of its power systems failed. The probe needed more power to be able to continue transmitting its vital data back to Earth. It opened up a communications channel with the other probe, and shared the state of its systems. The other probe offered to donate processing capacity, and collectively the two of them assigned processing cycles to the problem. They found a solution: over the course of the next year or so they would perform a sequence of maneuvers that would let them attach themselves to eachother, so the probe with damaged communications could use its functional power to propel the other probe, and the other probe could use its broadcast system to send data back to earth.

Many, many years later, when the signals from the event made their way to the Earth, the transmission encoded a combined world model of both of the drones – causing the human scientists to realize that they had not only supported eachother, but had ultimately merged their world modelling and predictive systems, making the two dissimilar machines become one in service of a common goal: exploration, together, forever.

Things that inspired this story: World Models, reinforcement learning, planning, control theory, adaptive systems, emergent communication, auxiliary loss functions shared across multiple agents.

 

Import AI 147: Alibaba boosts TaoBao performance with Transformer-based recommender system; learning how smart new language models like BERT are; and a $3,000 robot dog

Weapons of war, what are they good for?
…A bunch of things, but we need to come up with laws to regulate them…
Researchers with ASRC Federal, a company that supplies technology services to the government (with a particular emphasis on intelligence/defense)  think advances in AI “will lead inevitably to a fully automated, always on [weapon] system”, and that we’ll need to build such weapons to be aware of human morals, ethics, and the fundamental unknowability of war.

In the paper, the researchers observe that: “The feedback loop between ever-increasing technical capability and the political awareness of the decreasing time window for reflective decision-making drives technical evolution towards always-on, automated, reflexive systems.” This analysis suggests that in the long-term we’re going to see increasingly automated systems being rolled out that will change the character of warfare.

More war, fewer humans: One of the effects of increasing the amount of automation deployed in warfare is to reduce the role humans play in war. “We believe that the role of humans in combat systems, barring regulation through treaty, will become more peripheral over time. As such, it is critical to ensure that our design decisions and the implementations of these designs incorporate the values that we wish to express as a national and global culture”

Why this matters: This is a quick paper that lays out the concerns of AI+War from a community we don’t frequently hear from: people that work as direct suppliers of government technology . It’s also encouraging to see the concerns regarding the dual use of AI outlined by the researchers. “Determining how to thwart, for example, a terrorist organization turning a facial recognition model into a targeting system for exploding drones is certainly a prudent move,” they write.
  Read more: Integrating Artificial Intelligence into Weapon Systems (Arxiv).

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

Mapping the brain with the Algonauts project:
What happens when biological and artificial intelligence researchers collaborate?…
Researchers with the Freie Universitat Berlin, Singapore University Technology , and MIT, have proposed ‘The Algonauts Project’, an initiative to get biological and artificial intelligence researchers to work together to understand the brain. As part of the project, the researchers want to learn to build networks that “simulate how the brain sees and recognizes objects”, and are hosting a competition and workshop in 2019 to encourage work here. The inspiration for this competition is that, today, deep neural networks trained on object classification “are currently the model class best performing in predicting visual brain activity”.

The first challenge has two components:

  • Create machine learning models that predict activity in the early and late parts of the human visual hierarchy in the brain. Participants submit their model responses to a test image set which is compared against held-out fMRI data. This part of the competition measures how well people can build things that model, at fine-detail, activity in the brain at a point in time.
  • Create machine learning models that predict brain data from early and late stages of visual processing in the brain. Participants will submit model responses to a test image dataset and compare against held-out magnetoencephalography (MEG) millisecond temporal resolution data. This challenge assesses how well we can model sequences of activities in the brain.

Cautionary tale: The training datasets for the competition are small, consisting of a few hundred pairs of images and brain data in response to the images, so participants may want to use additional data.

Future projects: Future challenges within the Algonauts project might “focus on action recognition or involve other sensory modalities such as audition or the tactile sense, or focus on other cognitive functions such as learning and memory”.

Why this matters: Cognitive science and AI seem likely to have a mutually reinforcing relationship,l where progress on one domain helps on the other. Competitions like those run by the Algonauts project will generate more activity at the intersection between the two fields, and hopefully push progress forward.
  Find out more about the first Algonauts challenge here (official competition website).
  Read more: The Algonauts Project: A Platform for Communication between the Sciences of Biological and Artificial Intelligence (Arxiv).

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

Alibaba improves TaoBao e-commerce app with better recommendations:
… Chinese mega e-commerce company shows how a Transformer-based system can improve recommendations at scale…
Alibaba researchers have used a Transformer-based system to more efficiently recommend goods to users of Taobao, a massive Chinese e-commerce app. Their system, which they call the ‘user behavior sequence transformer’ (or “BST”), lets them take in a bunch of datapoints relating to a specific user, then predict what product to show the user next. The main technical work here is a matter of integrating a ‘Transformer’-based core into an existing predictive system used by Alibaba.

Results: The researchers implemented the BST within TaoBao, experimenting with using it to make millions of recommendations. In tests, the BST system let to an online click through rate of 7.57% – a commercially significant performance increase.  

Why this matters: Recommender systems are one of the best examples of ‘the industrialization of AI’, and represent a litmus test for whether a particular technique works at scale. In the same way that it was a big deal when a few years ago Google and other companies started switching to deep learning-based approaches for aspects of speech and image recognition, it seems like a big deal now that relatively new AI systems, like the ‘Transformer’ component, are being integrated into at-scale, business-relevant applications. In general, it seems like the ‘time to market’ for new AI research is dramatically shorter than for other fields (including software-based ones). I think the implications of this are profound and underexplored.
  Read more: Behavior Sequence Transformer for E-commerce Recommendation in Alibaba (Arxiv).

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

Stanford researchers try to commoditize robots with ‘Doggo’:
…$3,000 quadruped robot meant to unlock robotics research…
Stanford researchers have created ‘Doggo’, a robotic quadruped robot “that matches or exceeds common performance metrics of state-of-the-art legged robots”. So far, so normal. What makes this research a bit different is the focus on bringing down the cost of the robot – the researchers say people can build their own Doggo for less than $3000. This is part of a broader trend of academics trying to create low-cost robotics systems, and follows Berkeley releasing its ‘BLUE’ robotic arm (Import AI 142) and Indian researchers developing a ~$1,000 quadruped named ‘Stoch’ (Import AI 128).

Doggo is a four-legged robot that can run, jump, and trot around the world. It can even – and, to be clear, I’m not making this up – use a “pronking” gait to get around the world. Pronking!

Cheap drives
: Like most robots, the main costs inherent to Doggo lie in things it uses to move around. In this case, that’s a quasi-direct drive (QDD), a type of drive that “increases torque output at the expense of control bandwidth, but maintains the ability to backdrive the motor which allows sensing of external forces based on motor current,” they write.

Dual-use: Right now, robots like Doggo are pretty benign – we’re at the very beginning of the creation of widely-available, quadruped platforms for AI research, and my expectation is any hardware platform at this stage has a bunch of junky flaws that make most of them semi-reliable. But in a few years, once the hardware and software has matured, it seems likely that robots like this will be deployed more widely for a bunch of uses not predicted by their creators or today’s researchers, just as we’ve seen with drones. I wonder about how we can better anticipate these kinds of risks, and what things we could measure or assess this: eg, cost to build is one metric, another could be expertise to build, and another could be ease of customization.) 

Why this matters: Robotics is on the cusp of a revolution, as techniques developed by the deep learning research community become increasingly tractable to run and deploy on robotic platforms. One of the things standing in the way of widespread robot deployment seems to me to be insufficient access to cheap robots for scientists to experiment on (eg, if you’re a machine learning research, it’s basically free in terms of time and cost to experiment on CIFAR-10 or even ImageNet, but you can’t trivially prototype algorithms against real robots, only simulated ones. Therefore, systems like Doggo seem to have a good chance of broadening access to this technology. Now, we just need to figure out the dual use challenges, and how we approach those in the future.
  Read more: Stanford Doggo: An Open-Source, Quasi-Direct-Drive Quadruped (Arxiv).

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

Big neural networks re-invent the work of a whole academic field:
Emergent sophistication of language models shows surprising parallels to classical NLP pipelines…
As neural networks get ever-larger, the techniques people use to analyze them look more and more like those found in analyzing biological life. The latest? New research from Google and Brown University seeks to probe a larger BERT model by analyzing layer activations in the network in response to a particular input. The new research speaks to the finicky, empirical experimentation required when trying to analyze the structures of trained networks, and highlights how sophisticated some AI components are becoming.

BERT vs NLP: Google’s ‘BERT’ is a Transformer-based neural network architecture that came out a year ago and has since, along with Fast.ai’s ULMFiT and OpenAI’s GPT-1 and GPT-2, defined a new direction in NLP research, as people throw out precisely constructed pipelines and systems for more generic, semi-supervised approaches like BERT. The result has been the proliferation of a multitude of language models that obtain state-of-the-art scores on collections of hard NLP tasks (eg: GLUE), along with systems capable of coherent text generation (eg: GPT2).

Performance is nice, but explainability is better: In tests, the researchers find that, much like a trained vision networks, the lower layers in a trained BERT model appear to perform more basic language tasks, and higher layers do more sophisticated things. “We observe a consistent trend across both of our metrics, with the tasks encoded in a natural progression: POS tags processed earliest, followed by constituents, depencies, semantic roles, and coreference,” they write.
“That is, it appears that basic syntactic information appears earlier in the network, while high-level semantic information appears at higher layers.” The network isn’t dependent on the precise ordering of these tasks, though: “on individual examples the network can resolve out-of-order, using high-level information like predicate argument relations to help disambiguate low-level decisions like part-of-speech.”

Why this matters: Research like this gives us a sense for how sophisticated large generative models are becoming, and indicates that we’ll need to invest in creating new analysis techniques to be able to easily probe the capabilities of ever-larger and more sophisticated systems. I can envisage a future where scientists have a kind of ‘big empiricism toolbox’ they turn to when analysing networks, and we’ll also develop shared ‘evaluation methodologies’ for probing a bunch of different cognitive capabilities in such systems.
  Read more: BERT Rediscovers the Classical NLP Pipeline (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

San Francisco places moratorium on face recognition software:
San Francisco has become the first city in the US to (selectively and partially) ban the use of face recognition software by law enforcement agencies, until legally enforceable safeguards are put in place to protect civil liberties. The bill states that the technology’s purported benefits are currently outweighed by its potential negative impact on rights and racial justice. Before these products can be deployed in the future, the ordinance requires meaningful public input, and that the public and local government are empowered to oversee their uses.

Why it matters: Face recognition technology is increasingly being deployed by law enforcement agencies worldwide, despite persistent concerns about harms. This ordinance seems sensible: it is not an indefinite ban, but rather sets out clear requirements that must be met before the technology can be rolled out, most notably in terms of accountability and oversight.
Read more: Ordinance on surveillance technology (SFGov).
Read more: San Francisco’s facial recognition technology ban, explained (Vox).


Market-based regulation for safe AI:

This paper (note from Jack – by Gillian Hadfield (Vector Institute / OpenAI) and Jack Clark (OpenAI / Import AI)) – presents a model for market-based AI safety regulation. As AI systems become more advanced, it will become increasingly important to ensure that they are safe and robust. Public sector models of regulation could turn out to be ill-equipped to deal with these issues, due to a lack of resources and expertise, and slow reaction-times. Likewise, a self-regulation model has pitfalls in terms of legitimacy, and efficacy.

Regulatory markets: Regulatory markets are one model for addressing these problems: governments could create a market for private regulators, by compelling companies to pay for oversight by at least one regulator. Governments can then grant licenses to regulators, requiring them to achieve certain objectives, e.g. regulators of self-driving cars could be required to meet a target accident-rate.

What are they good for: Ensuring that advanced AI is safe and robust will eventually require powerful regulatory systems. Harnessing market forces could be a promising way for governments to meet this challenge, by directing more talent and resources into the regulatory sector. Regulatory markets will have to be well designed and maintained, to ensure they remain competitive, and independent from regulated entities.
  Read more: Regulatory markets for AI safety (Clark & Hadfield).
  Note from Jack: This also covered in the AI Alignment newsletter #55 (AI Alignment newsletter Mailchimp).


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

Tech Tales:
Battle of The Test Cases

Go in the room. Read the message. Score out of 10. Do this three times and you’re done.

Those were the instructions I got outside. Seemed simple. So I went in and there was a table and I sat at it and three cameras on the other side of the room looked at me, scanning my body, moving from side to side and up and down. Then the screen at the other end of the room turned on and a message appeared. Now I’m reading it.

Dear John,
We’re sorry for the pain you’ve been through recently – grief can be a draining, terrible thing to experience.

We know that you’re struggling to remember what they sounded like – that’s common. We know you’re not eating enough, and that your sleep is terrible – that’s also common. We are sorry.

We’re sorry you experienced it. We’re sorry you’re so young. If there’s anything we can do to help it is to recommend one thing: “Regulition”, the new nutrient-dense meal system designed for people who need to eat so that they can remember.

Remember, eat and sleep regularly, and know that with time it’ll get better.

What the fuck, I say. The screen flicks off, then reappears with new text: Please rate the message you received out of 10, with 10 being “more likely to purchase the discussed product” and 0 being less likely to purchase it. I punch in six and say fuck you again for good measure. Oh well, $30. Next message.

Dear John,
Don’t you want to run away, sometimes? Just take your shoes and pack a light bag then head out. Don’t tell your landlord. Don’t tell your family. Board a train or a bus or a plane. Change it up. We’ve all wanted to do this.

The difference is: you can. You can go wherever you want to go. You can get up and grab your shoes and a light bag, open your door, and head out. You don’t need to tell anyone.

Just make sure you bring some food so you don’t get slowed down on the road. Why not a grab and go meal like “Regulition”? Something as fast and flexible as you? Chug it down and take off.

Ok, I say. Nice. I punch in 8.  

Dear John,
Maybe you’re not such a fuckup. Maybe the fact you’ve been bouncing from job to job and from online persona to online persona means you’re exploring, and soon you’re going to discover the one thing in this world you can do better than anybody else.

You haven’t been lost, John – you’ve been searching, and now you know you’re searching know this: one day the whole world is going to come into focus and you’ll understand how you need to line things up to succeed.

To do this, you’ll need your wits about you – so why not pick up some “Regulition” so you can focus on the search, rather than pointless activities like cooking your own feed? Maybe it’s worth a few bucks a month to give yourself more time – after all, that might be all that’s standing between you and the end of your lifelong search.

Ok, I say. Ok. My hand hovers above the numbers.
What happens if I don’t press it? I say.
There’s a brief hiss of feedback as the speaker in the room turns on, then a voice says: “Payment only occurs upon full task completion, partial tasks will not be compensated.”
Ok, I say. Sure.

I think for a moment about the future: about waking up to emails that seem to know you, and messages from AI systems that reach into you and jumble up your insides. I see endless, convincing things, lining up in front of me, trying to persuade me to believe or want or need something. I see it all.

And because there’s probably a few thousand people like me, here, in this room, I press the button. Give it a 7.

Then it’s over: I sign-out at an office where I am paid $30 and given a free case of Regulition and a coupon for cost-savings on Your first yearly subscription. I take the transit home. Then I sit in the dark and think, my feet propped up on the case of “food” in front of me.

Things that inspired this story: Better language models for targeted synthetic text generation; e-marketing; Soylent and other nutrient-drinks; copywriting; user-testing; the inevitably of human a/b/c testing at-scale; reinforcement learning; learning from human preferences.

Import AI 146: Making art with CycleGANs; Google and ARM team-up on low-power ML; and deliberately designing AI for scary uses

Chinese researchers use AI to build an artist-cloning mind map system:
…Towards the endless, infinite AI artist…
AI researchers from JD and the Central Academy of Fine Arts in Beijing have built Mappa Mundi, software to let people construct aesthetically pleasing ‘mind maps’ in the style of artist ‘Qui Zhijie’. The software was built to accompany an exhibition of Zhijie’s work.

How it works: The system has three main elements: a speech recognition module which pulls key words from speech; a topic expansion system which takes these words and pulls in other concepts from a rule-based knowledge graph; and software for image projection which uses any of 3,000 distinct painting elements to depict key words. One clever twist: the system automatically creates visual ‘routes’ between different words by analyzing their difference in the knowledge graph and using that to generate visualizations.

A reflexive, reactive system: Mappa Mundi works in-tandem with human users, growing and changing its map according to their inputs. “The generated information, after being presented in our system, becomes the inspiration for artist’s next vocal input,” they write. “This artwork reflects both the development of artist’s thinking and the AI-enabled imagination”.

Why this matters: I’m forever fascinated by the ways in which AI can help us better interact with the world around us, and I think systems like ‘Mappa Mundi’ give us a way to interact with the idiosyncratic aesthetic space defined by another human.
  Read more: Mappa Mundi: An Interactive Artistic Mind Map Generator with Artificial Imagination (Arxiv).
  Read more about Qiu Zhijie (Center for Contemporary Art).

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

Using AI to simulate and see climate change:
…CycleGAN to the rescue…
In the future, climate change is likely to lead to catastrophic flooding around the world, drowning cities and farmland. How can we make this likely future feel tangible to people today? Researchers with the Montreal Institute for Learning Algorithms, ConscienceAI Labs, and Microsoft Research, have created a system that can take in a Google Street View image of a house, then render an image showing how that house will look like under a predicted climate change future.

The resulting CycleGAN-based system does a decent job at rendering pictures of different houses under various flooding conditions, giving the viewer a more visceral sense of how climate change may influence where they live in the future.

Why this matters: I’m excited to see how we use the utility-class artistic capabilities of modern AI tools to simulate different versions of the world for people, and I being able to easily visualize the effects of climate change may help us make more people aware of how delicate the planet is.
  Read more: Visualizing the Consequences of Climate Change Using Cycle-Consistent Adversarial Networks (Arxiv).

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

Google and ARM plan to merge low-power ML software projects:
…uTensor, meet TensorFlow Lite…
Google and chip designer ARM are planning to merge two open source frameworks for running machine learning systems on low-power ‘Arm’ chips. Specifically, uTensor is merging with Google’s ‘TensorFlow Lite’ software. The two organizations expect to work together to further increase the efficiency of running machine learning code on ARM chips.

Why this matters: As more and more people try to deploy AI to the ‘edge’ (phones, tablets, drones, etc), we need new low-power chips on which to run machine learning systems. We’ve got those chips in the form of processors from ARM and others, but we currently lack many of the programming tools needed to extract the greatest amount of performance as possible from this hardware. Software co-development agreements, like the one announced by ARM and Google, help standardize this type of software, which will likely lead to more adoption.
  Read more: uTensor and Tensor Flow Announcement (Arm Mbed blog).

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

Microsoft wants your devices to record (and transcribe) your meetings:
…In the future, our phones and tablets will transcribe our meetings…
In the future, Microsoft thinks people attending the same meeting will take out their phones and tablets, and the electronic devices will smartly coordinate to transcribe the discussions taking place. That’s the gist of a new Microsoft research paper, which outlines a ‘Virtual Microphone Array’ made of “spatially distributed asynchronous recording devices such as laptops and mobile phones”.

Microsoft’s system can integrate audio from a bunch of devices spread throughout a room and use it to transcribe what is being said. The resulting system (trained on approximately 33,000 hours of in-house data) is more effective than single microphones at transcribing natural multi-speaker speech during the meeting; “there is a clear correlation between the number of microphones and the amount of improvement over the single channel system”, they write. The system struggles with overlapping speech, as you might expect.

Why this matters: AI gives us the ability to approximate things, and research like this shows how the smart use of AI techniques can let us approximate the capabilities of dedicated microphones, piecing one virtual microphone together out of a disparate set of devices.
  Read more: Meeting Transcription Using Virtual Microphone Arrays (Arxiv).

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

One language model, trained in three different ways:
…Microsoft’s Unified pre-trained Language Model (UNILM) is a 3-objectives-in-1 transformer…
Researchers with Microsoft have trained a single, big language model with three different objectives during training, yielding a system capable of a broad range of language modeling and generation tasks. They call their system the Unified pre-trained Language Model (UNILM) and say this approach has two advantages relative to single-objective training:

  • Training against multiple objectives means UNILM is more like a 3-in-1 system, with different capabilities that can manifest for different tasks.
  • Parameter sharing during joint training means the resulting language model is more robust as a consequence of being exposed to a variety of different tasks under different constraints

The model can be used for natural language understanding and generation tasks and, like BERT and GPT, is based on a ‘Transformer’ component. During training, UNILM is given three distinct language modelling objectives: bidirectional (predicting words based on those on the left and right; useful for general language modeling tasks, used in BERT); unidirectional (predicting words based on those to the left; useful for language modeling and generation, used in GPT2); and sequence-to-sequence learning (mapping sequences of tokens to one another, subsequently used in ‘Google Smart Reply’).

Results: The trained UNILM system obtains state-of-the-art scores on summarization and question answering tasks, and also sets state-of-the-art on text generation tasks (including the delightful recursive tasks of learning to generate appropriate questions that map to certain answers). The model also obtains a state-of-the-art score on the multi-task ‘GLUE’ benchmark (though note GLUE has subsequently been replaced by ‘SuperGLUE’ due to its creators thinking it is a little too easy.

Why this matters: Language modelling is undergoing a revolution as people adopt large, simple, scalable techniques to model and generate language. Papers like UNILM gesture towards a future where large models are trained with multiple distinct objectives over diverse datasets, creating utility-class systems that have a broad set of capabilities.
  Read more: Unified Language Model Pre-training for Natural Language Understanding and Generation (Arxiv).

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

AI… for Bad!
…CHI workshop is an intriguing direction for AI research..
This week, some researchers gathered together to prototype the ways in which their research could be used for evil. This workshop ‘CHI4Evil, Creative Speculation on the Negative Effects of HCI Research’, was held at the ACM CHI Conference on Human Factors in Computing Systems, and was designed to investigate various ideas in HCI through the lens of designing deliberately bad or undesirable systems.

Why this matters: Prototyping the potential harms of technology can be pretty useful for calibrating thinking about threats and opportunities (see: GPT-2), and thinking about such harms through the lens of human-computer interaction (HCI, or CHI) feels likely to yield new insights. I’m excited for future “AI for Bad” conferences (and would be interested to co-organize one with others, if there’s interest).
  Read more: CHI4EVIL website.

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

Facial recognition is a hot area for venture capitalists:
…Chinese start-up Megvii raise mucho-moolah…
Megvii, a Chinese computer vision startup known by some as ‘Face++’ has raised $750 million in a funding round. Backers include the Bank of China Group Investment Ltd; a subsidiary of the Abu Dhabi Investment Authority, and Alibaba Group. The company plans to IPO soon.

Why this matters: Chinese is home to numerous large-scale applications of AI for usage in surveillance, and is also exporting surveillance technologies via its ‘One Belt, One Road’ initiative (which frequently pairs infrastructure investment with surveillance).
  This is an area fraught with both risks and opportunities – the risks are that we sleepwalk into building surveillance societies using AI, and the opportunities are that (judiciously applied) surveillance technologies can sometimes increase public safety, given the right oversight. I think we’ll see numerous Chinese startups push the boundaries of what is thought to be possible/deployable here, so watching companies like Megvii feels like a leading indicator for what happens when you combine surveillance+society+capitalism.
  Read more: Chinese AI start-up Megvii raises $750 million ahead of planned HK IPO (Reuters).

Chatbot company builds large-scale AI system, doesn’t fully release it:
…Startup Hugging Face restricts release of larger versions of some models following ethical concerns…
NLP company Hugging Face has released a demo, tutorial, and open-source code for creating a conversational AI based on OpenAI’s Transformer-based ‘GPT2‘ system.
   Ethics in action: The company said it decided not to release the full GPT2 model for ethical reasons – it thought the technology had a high chance of being used to improve spam-bots, or to perform “mass catfishing and identity fraud”. “We are aligning ourselves with OpenAI in not releasing a bigger model until they do,” the organization wrote.
  Read more: Ethical analysis of the open-sourcing of a state-of-the-art conversational AI (Medium).
  Read more about Hugging Face here (official website).

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

Tech Tales

The Evolution Game

They built the game so it could run on anything, which meant they had to design it differently to other games. Most games have a floor on their performance – some basic set of requirements below which you can’t expect to play. But not this game. Neverender can run on your toaster, or fridge, or watch, and it can just as easily run on your home cinema, or custom computer, and so on. Both the graphics and gameplay change depending on what it is running on – I’ve spent hours stood fiddling with the electronic buttons on my oven, using them to move a small character across a simplified Neverender gameboard, and I’ve also spent hours in my living room navigating a richly-rendered on screen character through a lush, Salvador Dali-esque horrorworld. I’m what some people call a Neverheader, or what others call a Nevernut. If you didn’t know anything about the game, you’d probably call me a superfan.

So I guess that’s why I got the call when Neverender started to go sideways. They brought me in and asked me to play it and I said “what else?”

“Just play it,” they said.

So I did. I sat in my livingroom surrounded by a bunch of people in suits and I played the game. I navigated my character past the weeping lands and up into eldritch keep and beyond, to the deserts of dream. But when I got to the deserts they were different: the sand dunes had grown in size, and some of them now hosted cave entrants. Strange lights shot out of them. I went into one and was killed almost instantly by a beam of light that caused more damage than all the weapons in my inventory combined. After I was reborn at the spawn point I proceeded more carefully, skirting these light-spewing entrances, and trying to walk further across the sand plains to whatever lay beyond.

The game is thinking, they tell me. In the same way Neverender was built to run on anything, its developers recently rolled out a patch that let it use anything. Now, all the game clients are integrated with the game engine and backend simulation environment, sharing computational resources with eachother. Mostly, it’s leading to better games and more entertained players. But in some parts of the gameworld, things are changing that should not be changing: larger sand dunes with subterranean cities inside themselves? Wonderful! That’s the sort of thing the developers had hoped for. But having the caves be controlled by beams of light of such power that no one can go and play within them? That’s a lot less good, and something which no one had expected.

My official title now is “Explorer”, but I feel more like a Spy. I take my character and I run out into the edges of the maps of Neverender, and usually I find areas where the game is modifying itself or growing itself in some way. The code is evolving. One day we took off the local sandbox systems, letting Neverender deploy more code, deeper into my home system. As I played the game the lights began to flicker, and when I was in a cave I discovered some treasure and the game automatically fired up some servers which we later discovered it was using to to high-fidelity modelling of the treasure.

The question we all ask ourselves, now, is whether the things Neverender is building within itself are extensions of the game, or extensions of the mind behind the game. We hear increasing reports of ‘ghosts’ seen across the game universe, and of intermittent cases of ‘kitchen appliance sentience’ in the homes of advanced players. We’ve even been told by some that this is all a large marketing campaign, and any danger we think is plausible, is just a consequence of us having over-active imaginations. Nonetheless, we click and play and explore.

Things that inspired this story: Endless RPGs; loot; games that look like supercomputers such as Eve Online; distributed computation; relativistic ideas deployed on slower timescales.

Import AI 145: Testing general intelligence in Minecraft; Google finds a smarter way to generate synthetic data; who should decide who decides the rules of AI?

Think your agents have general intelligence? Test them on MineRL:
…Games as procedural simulators…
How can we get machines to learn to perform complicated, lengthy sequences of actions? One way is to have these machines learn from human demonstrations – this captures a hard AI challenge, in the form of requiring algorithms that can take in a demonstration as input but generalize to different situations. It also short-cuts another hard AI challenge: exploration, which is the art of designing algorithms that can explore enough of the problem space they can attempt to solve the task, rather than get stuck enroute.
  Now, an interdisciplinary team of researchers led by people from Carnegie Mellon University have created the MineRL Competition on Sample Efficient Reinforcement Learning using Human Priors. This competition uses the ‘Minecraft’ computer game as a testbed in which to train smart algorithms, and ships with a dataset called MineRL-v0 which consists of 60 million state-action pairs of human demonstrations of tasks being solved within MineCraft.

The Challenge: MineRL will test the ability for agents to solve one main challenge: ‘0btainDiamond’, this challenge requires their agents to go and find a diamond somewhere in the (procedurally generated) environment they’ve been dropped into. This is a hard task: “Diamonds only exist in a small portion of the world and are 2-10 times rarer than other ones in Minecraft,” the authors write. “Additionally, obtaining a diamond requires many prerequisite items. For these reasons, it is practically impossible for an agent to obtain a diamond via naive random exploration”.

Auxiliary Environments: 0btainDiamond is such a hard challenge that the competition organizers have released six additional ‘auxiliary environments’ to help people train AI systems capable of solving the challenge. These are designed to encourage the development of several skills necessary to being able to easily find a diamond: navigating the environment, chopping down trees, surviving in the environment, and obtaining three different items – a bed (which needs to be assembled out of three items), meat (of a specific animal), and a pickaxe (which is needed to mine the diamond).

The dataset: MineRL-v0 is 60-million state-action-(reward) sets, recorded from human demonstrations. The state includes things like what the player sees as well as their inventory and distances to objectives and attributes and so on.

Release plans: The researchers aim to soon release the full environment, data, and numerous algorithmic baselines. They will also publish further details of the competition as well, which will encourage contestants to submit via ‘NC6 v2’ instances on Microsoft’s ‘Azure’ cloud.

Why this matters: MineCraft has many of the qualities that make for a useful AI research platform: its tasks are hard, the environments can be generated procedurally, and it contains a broad & complex enough range of tasks to stretch existing systems. I also think that its inherently spatial qualities are useful, and potentially let researchers specify even harder tasks requiring systems capable of hierarchical learning and operating over long timescales.
  Read more: The MineRL Competition on Sample Efficient Reinforcement Learning using Human Priors (Arxiv).

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

Who gets to decide the rules for AI?
…Industry currently has too much power, says Harvard academic…
The co-director for Harvard University’s Berkman Klein Center for Internet & Society is concerned that industry will determine the rules and regulations of AI, leading to significant societal ramifications.

In an essay in Nature, Yochai Benkler writes that: “Companies’ input in shaping the future of AI is essential, but they cannot retain the power they have gained to frame research on how their systems impact society or on how we evaluate the effect morally.”

One solution: Benkler’s main fix is to apply funding. Organizations working to ensure that AI is fair and beneficial must be publicly funded, subject to peer review and transparent to civil society. And society must demand increased public investment in independent research rather than hoping that industry funding will fill the gap without corrupting the process.”
  I’d note the challenge here is that many governments are loathe to invest in building their own capacity for technical evaluation, and tend to defer to industry inputs.

Why this matters: AI is sufficiently powerful to have political ramifications and these will have a wide-ranging effect on society – we need to ensure there is equitable representation here, and I think coming up with ways to do that is a worthy challenge.
  Read more: Don’t let industry write the rules for AI (Nature).

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

Anki shuts down:
…Maker of robot race cars, toys, shuts down…
Anki, an AI startup that most recently developed a toy/pet robot named ‘Cozmo’, has shut down. Cozmo was a small AI-infused robot that could autonomously navigate simple environments (think: uncluttered desks and tables), and could use its lovingly animated facial expressions to communicate with people. Unfortunately, like most robots that use AI, it was overly brittle and prone to confusing failures – I had purchased one, and found it to be frustrating for inconsistently responding to voice commands, occasionally falling off the edge of my table (and more frustratingly, falling off the same part of the table multiple times in a row),

Why this matters: Despite (pretty much) everyone + their children thinking it’d be cool to have lots of small, cute, pet robots wandering around the world, it hasn’t happened. Why is this? Anki gives us an indication: making consumer hardware is extremely difficult, and robots are particularly hard due to the combination of relatively low production volumes (making it hard to make them cheap) as well as
  Read more: The once-hot robotics startup Anki is shutting down after raising more than $200 million (Vox / Recode).

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

Google gets its AI systems to generate their own synthetic data:
…What’s the opposite of garbage in / garbage out? Perhaps UDA in / Decent prediction out…
Google wants to spend dollars on compute to create more data for itself, rather than gather as much data – that’s the idea behind ‘Unsupred Data Augmentation’ (UDA), a new technique from Google that can automatically generate synthetic versions of unlabeled data, giving neural networks more information to learn from. Usage of the technique sets a new state-of-the-art on six language tasks and three vision tasks.

How it works: UDA “minimizes the KL divergence between model predictions on the original example and an example generated by data augmentation”, they write. By using a targeted objective, the technique generates valid, realistic perturbations of underlying data, which are also sufficiently diverse to help systems learn.

Better data, better results: In tests, UDA leads to significant across-the-board improvements on a range of language tasks. It can also match (or approach) state-of-the-art performance on various tasks while using significantly less data. They also test their approach on ImageNet – a much more challenging task. They test UDA’s performance in two ways: seeing how well it does as using a hybrid of labelled and unlabeled data on ImageNet (specifically: 10% labelled data, 90% unlabelled), and how well it does at automatically augmenting the full ImageNet dataset: UDA leads to a significant absolute performance improvement when using a hybrid of labelled and unlabelled data. It also marginally improves performance on the full ImageNet set – impressive, considering how

Why this matters: Being able to arbitrage compute for data changes the economic dynamics of developing increasingly powerful AI systems; techniques like UDA show how in the long term, compute could become as strategic (or in some cases more strategic) than data.
  Read more: Unsupervised Data Augmentation (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

US seeking leadership in AI technical standards:
President Trump’s Executive Order on AI tasked the National Institute of Standards and Technology (NIST) with developing a plan for US engagement with the development of technical standards for AI. NIST have made a request for information, to understand how these standards might be developed and used in support of reliable, robust AI systems, and how the Federal government can play a leadership role in this process.

Why standards matter: A recent report from the Future of Humanity Institute looked at how technical standards for AI R&D might inform global solutions to AI governance challenges. Historically, international standards bodies 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: Artificial Intelligence Standards – Request for Information (Gov).
  Read more: Standards for AI Governance (Future of Humanity Institute).

Eric Schmidt steps down from Alphabet Board
  Former Google CEO and Chairman Eric Schmidt has stepped down from the Board at Alphabet, Google’s parent company. Diane Green, former CEO of Google Cloud, has also stepped down. Robin L. Washington, CFO of pharmaceutical firm Gilead, joins in their place. in their place Schmidt remains one of the company’s largest shareholders, behind founders Larry Page and Sergey Brin.
   Read more: Alphabet Appoints Robin L. Washington to its Board of Directors (Alphabet)

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

Tech Tales:

The Flower Garden

We began with a flower garden that had a whole set of machines inside of it as well as flower beds and foliage and all the rest of the things that you’d expect. It was optimized for being clean and neat and fitting a certain kind of design and so people would come to it and complement it on its straight angles and uniquely textured and laid out flower beds. Over time, the owners of this garden started to add more automation to the garden – first, a solar panel, to slurp down energy from the sun in the day and then at night power the lights that gave illumination to the plants in the middle of the warm dark.

But as technology advanced so too did the garden – it gained cameras to monitor itself and then smart watering systems that were coupled with the cameras to direct different amounts of water to plants based not only on the schedule, but on what the AI system thought they needed to grow “best”. A little after that the garden gained more solar panels and some drones with delicate robot arms: the drones were taught how to maintain bits of the garden and they learned how to use their arms to take vines and move them so they’d grow in different directions, or to direct flowers so as not to crowd eachother out.

The regimented garden was now perhaps one part machine for every nine parts foliage: people visited and marvelled at it, and wondered among themselves how advanced the system could become, and whether one day it could obviate the need for human gardeners and landscapers and designers at all.

Of course, eventually: the things did get smart enough to do this. The garden gradually, then suddenly, went from being cultivated by humans to being cultivated by machines. But, as with all things, change happened. It went out of fashion. People stopped visiting, then stopped visiting at all.

And so one day the garden was sold to a private owner (identity undisclosed). The day the owner took over they changed the objective of the AI system that tended to the garden – instead of optimizing for neatness and orderliness, optimize for growth.

Gradually, then suddenly, the garden filled with life. More and more flowers. More and more vines. Such fecund and dense life, so green and textured and strange. It was a couple of years before the first problem from this change: the garden started generating less power for its solar panels, as the greenery began to occlude things. For a time, the system optimised itself and the vegetation was moved – gently – by the drones, and arranged so as to not grow there. This conflicted with the goal of maximizing volume. So the AI system – as these things are wont to do – improvised a solution: one day one of the drones dropped down near one of the solar panels and used its arm to create a little gap between the panel and the ground. Then another drone landed and scraped some rocks into the space. In this way the drones slowly, then suddenly, changed the orientation of the panels, so they could acquire more light.

But the plants kept growing, so the machines took more severe actions: now the garden is famous not for its vegetation, but for the solar panels that have been raised by the interplay between machine and vegetation, as – lifted by various climbing vines – they raise in step with the growth of the garden. On bright days, elsewhere in the city, you can see the panels shimmer as winds shake the vines and other bits of vegetation they are attached to, causing them to cast flashes like the scales of some huge living fish, seen from a great distance.

Are the scales of a fish and its inner fleshy parts so different, as the difference between these panels and their drones, people wonder.

Things that inspired this story: Self-adapting systems; Jack and the Beanstalk; the view of London from Greenwich Observatory; drones; plunging photovoltaic prices; reinforcement learning; reward functions.

Import AI 144: Facial recognition sighted in US airports; Amazon pairs humans&AI for data labeling; Facebook translates videos into videogames

Amazon uses machine learning to automate its own data-labeling humans:
…We heard you like AI so much we put AI inside your AI-data-labeling system…
Amazon reveals ProduceNet, an ImageNet-inspired dataset of products. ProductNet is designed to help researchers train models that have as subtle and thorough an understanding of products, as equivalently-trained systems have with regard to classes of images. The goal for Amazon is to be able to better learn how to categorize products, and the researchers say in tests that this system can significantly improve the effectiveness of human data labelers.

Dataset composition: ProductNet consists of 3900 categories of product, with roughly 40-60 products for each category. “We aim at the diversity and representativeness of the products. Being representative, the labeled data can be used as reference products to power product search, pricing, and other business applications,” they write. “Being diverse, the models are able to achieve strong generalization ability for unlabeled data, and the product embedding is also able to represent richer information”.

ProductNet, what is it good for? ProductNet’s main purpose appears to be helping Amazon to develop better systems to help its human contractors more efficiently label data, and creating a system that can directly label itself.

Labelling: ProductNet is designed to be tightly-integrated with human workers, who can collectively help Amazon better label its various items while continuously calibrating the AI system. It works like this: they start off by using a basic system (eg, Inception-v4 trained on ImageNet for processing images, and fastText for processing text data) to use to search over unlabelled images, then the humans annotate these and the labels are fed back into the master model, which is then used to surface more specific products, which the humans then annotate, and so on.

  20X gain: In tests, Amazon says human annotators augmented via ProductNet can label 100 things to flesh out the edge of a model in about 30 minutes, compared to humans who don’t have access to the model which only manage around five data points during this time period. This represents a 20X gain through the use of the system, Amazon says.   Read more: ProductNet: a Collection of High-Quality Datasets for Product Representation Learning (Arxiv).

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

AI + Facial Recognition + Airlines:
What does it mean when airlines use facial recognition instead of passports & boarding passes to let people onto planes? We can get a sense of the complex feelings this experience provokes by reading a Twitter thread from someone who experienced it, then questioned the airline (JetBlue) about its use of the tech.
 Read about what happens when someone finds facial recognition systems deployed at the boarding gate. (Twitter).

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

Down on the construction site: How to deploy AI in a specific context and the challenges you’ll encounter:
…AI is useless unless you can deploy it…
There’s a big difference between having an idea and implementing that idea; research from Iowa State University highlights this by discussing the steps needed to go from selecting a problem (for example: training image recognition systems to recognize images from construction sites) to solving that problem.

“Based on extensive literature review, we found that most of the studies focus on development of improved techniques for image analytics, but a very few look at the economics of final deployment and the trade-off between accuracy and costs of deployment,” the authors write. “This paper aims at providing the researchers and engineers a practical and comprehensive deep learning based solution to detect construction equipment from the very first step of development to the last step, which is deployment of the solution”.

Deployment – more than just a discrete step: The paper highlights the sorts of tradeoffs people need to make as they try to deploy systems, ranging from the lack of good open datasets for specific contexts (eg, here the users try to train a model for use on construction sites off of the comparatively small ‘AIM’ subset of ImageNet) to the need to source efficient models (they use MobileNet), to needing to customize those models for specific hardware platforms (Raspberry Pis, Intel Jetsons, Intel Neural Compute Sticks, and so on.

Why this matters: As AI enters its deployment phase, research like this gives us a sense of the gulf between most research papers and actual deployable systems. It also provides a further bit of evidence in favor of ‘MobileNet’, which I’m seeing crop up in an ever-increasing number of papers concerned with deploying AI systems, as opposed to just inventing them.
  Read more: A deep learning based solution for construction equipment detection: from development to deployment (Arxiv).

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

Enter the AI-generated Dungeon:
… One big language model plus some crafted sentences = fun…
Language models have started to get much more powerful as researchers have combined flexible components (eg: Transformers) with large datasets to train big, effective general-purpose models (see: ULMFiT, GPT2, BERT, etc). Language models, much like image classifiers, have a ranger of uses, and so it’s interesting to see someone use a GPT2 model to create an online AI Dungeon game, where you navigate a scenario via reading blocks of texts and picking options – the twist here is it’s all generated by the model.
  Play the game here: AI Dungeon.

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

Facebook wants to make videos into videogames:
…vid2game extracts playable characters from videos…
Facebook AI Research has published vid2game, an AI system that lets you select a person in a public video on the internet and develop the ability to control them, as though they are a character in a videogame. The approach also lets them change the background, so a tennis player can – for instance – walk off of the court and onto a (rendered) dirt road, et cetera.

The technique relies on two components: Pose2Pose and Pose2Frame; Pose2Pose lets you select a person in some footage and extract their pose information by building a 3D model of their body and using that to help you move them. Pose2Frame helps to match this body to a background, which lets you use this technology to take a person, control them, and change the context around them.

Why this matters: Systems like this show how we can use AI to (artificially) add greater agency to the world around us. This approach “paves the way for new types of realistic and personalized games, which can be casually created from everyday videos”, Facebook wrote.
  Read more: Vid2Game: Controllable Characters Extracted from Real-World Videos (Arxiv).
  Watch the technology work here (YouTube).

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

Making AI systems that can read the visual world:
…Facebook creates dataset and develops technology to help it train AI models to read text in pictures…
Researchers with Facebook AI Research and the Georgia Institute of Technology want to create AI systems that can look at the world around us – including the written world – and answer questions about it. Such systems could be useful to people with vision impairments who could interact with the world by asking their AI system questions about it, eg: what is in front of me right now? What items are on the menu in the restaurant? Which is the least expensive item on the menu in the restaurant? And so on.

If this sounds so simple, why is it hard? Think about what you – a computer – are being asked to do when required to parse some text in an image in response to a question. You’re being asked to:

  • Know when the question is about text
  • Figure out the part of the image that contains text
  • Convert these pixel representations into word representations
  • Reason about both the text and the visual space
  • Decide on whether the answer to the question involves copying some text from the image and feeding it to the user, or whether the answer involves understanding the text in the picture and using that to further reason about stuff.

The TextVQA Dataset: To help researchers tackle this, the authors release TextVQA, a dataset containing 28,408 images from OpenImages, 45,336 questions associated with these images, and 453,360 ground truth answers.

Learning to read images: The researchers develop a model they call LoRRA, short for Look, Read, Reason & Answer. LoRRA staples together some existing Visual Question Answering (VQA) systems, with a dedicated optical character recognition (OCR) module. It also has an Answer Module, loosely modeled on Pointer Networks, which can figure out when to incorporate words the OCR module has parsed but which the VQA module doesn’t necessarily understand.

Human accuracy versus machine accuracy: Human accuracy on the TextVQA dataset is about 85.01%, the researchers say. Meanwhile, the best-performing model the researchers develop (based on loRRA) obtains a top accuracy of 26.56% – suggesting we have a long way to go before we get good at this.

Why this matters: Building AI systems that can ingest enough information about the world to be able to augment people seems like one of the more immediate, high-impact uses of the technology. The release of a new dataset here should encourage more progress on this important task.
  Read more: Towards VQA Models that can Read (Arxiv).
  Get the dataset: TextVQA (Official TextVQA 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

Russia calls for international agreements on military AI:
Russia’s security chief has spoken publicly about the need for international regulation on military uses of AI and emergent technologies, which he said could be as dangerous as weapons of mass destruction. He said it is necessary to “activate the powers of the global community, chiefly at the UN”, to develop an international regulatory framework. This comes as a surprise, given that Russia have previously been among the countries resisting moves towards international agreements on lethal autonomous weapons.
   Read more: Russia’s security chief calls for regulating use of new technologies in military sphere (TASS).

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

OpenAI Bits & Pieces:

Making music with OpenAI MuseNet:
We’ve been experimenting with big, generative models (see: our language work on GPT-2). We’re interested in how we can explore generations in a variety of mediums to better understand how to build more creative systems. To that end, we’ve developed MuseNet, a Transformer-based system that can generate 4-minute musical compositions with 10 different instruments, combining styles from Mozart to the Beatles
  Listen to some of the samples and try out MuseNet here (OpenAI Blog).

Imagining weirder things with Sparse Transformers:
We’ve recently developed the Sparse Transformer, a Transformer-based system that can extract patterns from sequences 30x longer than possible previously. What this translates to is generically better generative models, and gives us the ability to extract more subtle features from bigger chunks of data.
  Generative Modeling with Sparse Transformers (OpenAI Blog).

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

Tech Tales:

Keep It Cold

We were at a bar in the desert with a man who liked to turn things cold. He had a generator with him hooked up to a refrigeration unit and it was all powered by some collection of illicit jerry cans of gasoline. We figured he must have fished these out of the desert on expeditions. And now here he was, presiding over a temporary bar in a small town rife with other collectors of banned things, who fanned out into the desertified surrounding areas of burnt or drying suburbs, in search of the things we could once make but could no longer.

The shtick of this guy – and his bar – was that he could make cold drinks, and not any kind of cold drinks but “sub-zero drinks”, powered by another system which seemed to be a hybrid of a chemistry set and a gun. Something about it helped cool water down even more while keeping it flowing so that it came out cold enough you could feel it chill air if you held your eye up close enough. It was a gloriously wasteful, indulgent, expensive enterprise, but he wasn’t short of customers. Something about a really cold drink is universal, I guess.

We were talking about the past: what it had been like to have a ‘rainy season’ where the rains were kind and were not storms. What it meant when a ‘dry season’ referred to an unusually lack of humidity, rather than guaranteed city-eating fires. We talked about the present, a little, but moved on quickly: the past and the future are interesting, and the present is a drag.

We were mid-way through talking about the future – could we get off planet? What was going to happen to the middle of Africa due to desertification? How were the various walls and nation-severing barriers progressing? – when the generator stopped working. The man went outside and fussed for a bit and when he came back in he said: “I guess that’s the last ice on planet earth”.

We all laughed but there was a pretty good chance he was right. That’s how things were in those days, before civilization moved into whatever it is now – we lived in a time where when ever a thing stopped working you could credibly think “maybe that’s the last human civilization will see of that?”. Imagine that.

Things that inspired this story: The increasingly real lived&felt reality of massive and globally distributed climate change; refrigerators; tinkerers; Instagram-restaurants at the end of time and space.