Import AI 169: Multi-task testing for vision; floating drones could build communications networks in the future; medical tech gets FDA approval

PyTorch gets smarter on mobile devices:
…1.3 update adds efficiency-increasing experimental features…
PyTorch, an AI programming framework that integrates nicely into the widely-used Python language, has got into version 1.3. The latest update for the software includes features for making it easier to train models with lower-precision, and also to deploy them onto mobile devices with limited computational budgets. Along with tools for making AI systems developed within PyTorch more interpretable. 

Hardware support: AI frameworks are part of a broader, competitive landscape in AI development, and hardware/cloud support is where we can look for signs of success of a given framework. Therefore, it seems promising for PyTorch’s prospects that it is now supported by Google’s custom “TPU” processors, as well as being directly supported within Alibaba’s cloud. 

Why this matters: Programming languages, much like spoken languages, define the cultural context in which technology is produced. Languages are also tools of power in themselves – it’s no coincidence that PyTorch’s biggest backer is Facebook and the framework PyTorch is seeking to dethrone is TensorFlow; successful frameworks generate other strategic advantages for the company’s that develop them (see: TensorFlow being written with some specific components that support TPUs, etc).
   Read more: PyTorch 1.3 adds mobile, privacy, quantization, and named tensors (official PyTorch website)

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

Could drones let us build airborne communications networks?
…”Don’t be alarmed, citizens, we are automatically dispatching the communication drones to restore service”…
Taiwanese researchers envision a future where drones are used as flying base stations, providing communications and surveillance services to dense areas. Getting there is going to be a challenge, as multiple technologies need to be matured for such a technology to be possible, they write in a position paper. But if we’re able to surmount them, the advantages could be profound, giving civilizations the ability to create what can literally be described as a ‘smart cloud’ (of drones!) at will. 

What stands between us and a glorious drone future? The researchers think there are five challenges that stand between us and a glorious, capable drone future. These include:

  • Long-term placement: How can we build drones that can hover for long enough periods of time they could serve useful purposes for communications networks?
  • Crowd estimation: Can we integrate computer vision tech into our drones so they can automatically analyze crowds of people around them? (The answer here is ‘yes’ but some of the technologies are still a bit juvenile. See Import AI #167).
  • 3D placement: Where do you stick the drone to optimize for reliable communications?
  • Adaptive 3D placement: Can you automatically move the drone to new locations in 3D space according to data from another source? (E.g., can you predict where crowds are going to assemble and can you proactively move the drone there ahead of time?)
  • Smart back-haul: How do you optimize communications between your drones and their base stations?

Why this matters: Have you ever looked at the sky? It’s big! There’s room to do a lot of stuff in it! And with the recent advances in drone affordability and intelligence, we can expect our skies to soon gain a ton of drones for a variety of different purposes. I think it’d be a nice achievement for human civilization if we can use drones to provide adaptive infrastructure, especially after natural disasters; papers like this get us closer to that future.
   Read more: Communications and Networking Technologies for Intelligent Drone Cruisers (Arxiv)

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

Testing robots with… weights, hockey sticks, and giraffes?
Ghost Robotics, a startup making a small quadruped robot, showed off the stability of their machine recently by chucking a weight at it, knocking it off balance. Check out this tweet to see a short video of the robot nimbly recovering after being hit. 

Robots & perturbations: This video makes me think of all the different ways researchers are fooling around with robots these days, and it reminds me of Boston Dynamics pushing one of its robots over using a hockey stick, and in another video forcing its ‘Spot’ quadruped to slip on a banana skin. Even OpenAI (where I work) got into the action recently, with a “plush giraffe perturbation” that it applied to a robot hand trying to manipulate a Rubiks cube. 

Why this matters: I think these sorts of demonstrations give us a visceral sense of progress in robotics. What happens in a few decades when semi-sentient AI systems look at all of this violent human on robot content – how will they feel, I wonder?
   Check out the Ghost Robotics video here (Ghost Robotics Twitter).
   Watch OpenAI’s Plush Giraffe Perturbation here (OpenAI co-founder Greg Brockman’s Twitter account).

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

Tencent & Mirriad team-up to insert adverts into existing videos:
…Of all the gin-joints in all the towns in all the world, she walks into mine [WITH A BOTTLE OF GORDON’S GIN]…
Tencent has partnered with AI startup Mirriad to use AI to insert advertisements into existing media, like TV shows and films. In other words: look forward to seeing a “gin-joint” in Casablanca full of prominent “Gordon’s” gin bottles, or perhaps a Coca-Cola logo emblazoned on the side of a truck in Terminator. Who knows! “With Mirriad’s API, the integration will be fully automated with ease and speed to ultimately transform the way advertisers engage with their target audiences in content”. Mirriad as a tech-heavy company, claiming to have 29 patents and/or patents pending, according to its website. “We create ad inventory where none existed, offering a new revenue stream from existing assets,” the company says

Why this matters: Once people start making a ton of money via AI-infused businesses, then we can expect more investment in AI, which will lead to further use cases, which will further proliferate AI into society (for better or for worse). Deals like this show just how rapidly various machine learning techniques have matured, yielding new companies. It also shows that it’s getting really cheap to edit recorded reality.
   Read more: Mirriad Partners With Tencent, One of the World’s Largest Video Platforms, to Reach Huge Entertainment Audiences with Branded Content Solution (PRNewsWire)

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

Deepfakes are helping people to fake everything:
…What links UK PM Boris Johnson to a synthetic image? Read on to find out…
In the past few years, people have started using synthetic images of people to help them carry out nefarious activities. Earlier this year, for instance, the Associated Press reported on a LinkedIn profile allegedly used by a spy that had a fake identity replete with a synthetic face for a profile picture. As the technology has become cheaper, easier to access, and more well known, more people have been using it for nefarious purposes. The latest? Allegations that Hacker House, a company run by Jennifer Arcuri (a startup executive who has been connected to UK PM Boris Johnson), may not be entirely real. The evidence? It seems like at least one person – “Annie Tacker” – connected to the company is actually a LinkedIn profile with a synthetic image attached and little else, according to sleuthing from journalist Phil Kemp. AI analysis startup Deeptrace Labs backed this up, saying on Twitter that the headshot “shows telltale marks of GAN headshots synthesis” (specifically, they think the image was produced with StyleGAN). 

Why this matters: Reality, at least online reality, is becoming trivial to manipulate. At the same time, after several millions years evolving to believe what comes in via our vision stream, people are pretty susceptible to sufficiently good synthetic propaganda. Cases like this illustrate how contemporary AI systems are rapidly making their way into society, yielding changes in how people believe and, allegedly, scam. Expect more.
   Read more: Phil Kemp’s twitter thread // Deeptrace Labs’ twitter post.
   Read more: Experts: Spy used AI-generated face to connect with targets (Associated Press).

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

Computer, ENHANCE this brain image! AI tech gets FDA approval:
…Deep learning-based denoising and resolution enhancement – now approved for medical use by regulators…
Subtle Medical, an AI startup that uses deep learning-based technology to manipulate medical images to aid better diagnosis, has received clearance from the U.S. Food and Drug Administration to sell its ‘SubtleMR’ product. SubtleMR is “image processing software that uses denoising and resolution enhancement to improve image quality,” according to a press release from Subtle Medical. The technology is being piloted in several university hospitals and imaging centers. Subtle has published several research papers that use approaches like generative adversarial networks for medical work.

The FDA is getting faster at approving AI tools: The USA’s Food and Drug Administration (FDA) has been speeding up the rate at which it approves products that incorporate AI. This year, the agency announced plans to “consider a new regulatory framework specifically tailored to promote the development of safe and effective medical devices that use advanced artificial intelligence algorithms”. As part of this, the agency also released a white paper describing this framework. 

Why this matters: So far, many of the most visible uses of AI technology have been in consumer apps (think: face filters for SnapChat), surveillance (including state-level surveillance initiatives), and ambitious-but-yet-to-pan-out projects like self-driving cars. My intuition is people are going to be significantly more receptive to AI if it starts showing up in medicine to help cure more people (ideally at lower costs).
   Read more: Subtle Medical Receives FDA 510(k) Clearance for AI-Powered SubtleMR(™) (PR Newswire).
   Check out some of Subtle Medical’s publications here (official Subtle Medical website).
   Read more: Statement from FDA Commissioner Scott Gottlieb, M.D. on steps toward a new, tailored review framework for artificial intelligence-based medical devices (FDA).
   Check out the FDA whitepaper: Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning  (AI/ML)-Based Software as a Medical Device (SaMD) – Discussion Paper and Request for Feedback (Regulations.gov).

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

Germany increases funding for AI research:
…National research group creates strategic funding for AI…
The Deutsche Forschungsgemeinschaft (DFG), a German research funding organization, says it has earmarked 90 million Euros ($100 million) for the creation of new AI research groups in the country. The funding is designed in particular to create grants that give promising young scientists focused on AI significant grants to increase their autonomy. Proposals are due to be made in 2019 and funding decisions will be made as early as the beginning of 2020, the DFG said. 

Why this matters: Around the world, governments are waking up to the strategic importance of AI and are increasing their funding in response. Looked at in isolation, this German funding isn’t a huge deal, but given that most countries in the world are making strategic investments of similar sizes (at minimum!), the aggregate effect is going to be quite large.
   Read more in the official (German) press release (DFG website).

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

Are we really making progress in computer vision? The VTAB test might help us find out:
…New multi-task benchmark aims to do for vision what GLUE did for NLP…
In recent years, computers have got really good at image recognition – so good in fact that we need to move from looking at single benchmarks, like ImageNet classification scores, to suites of test that run one AI system through a multitude of evaluations. That’s the idea behind the Visual Task Adaptation Benchmark (VTAB), a testing suite developed by researchers at Google.

What is VTAB?
“VTAB is based on a single principle: a better algorithm is one that solves a diverse set [of] previously unseen tasks with fewest possible labels,” the researchers write. “The focus on sample complexity reflects our belief that learning with few labels is the key objective of representation learning.” 

Tasks and categories: VTAB contains 19 tasks split across three categories:

  • “Natural” category: This includes classification tasks over widely-used datasets such as Caltech101, Flowers102, and SVHN
  • “Specialized” category: This uses datasets that contain images captured with specialized equipment, split across satellite and medical imagery. 
  • “Structured” category: This tests how well a system understands the structure of a scene and evaluates systems according to how well they can count objects in pictures, or estimate the depth of various visual scenes. 

How good are existing models at VTAB? The researchers test 16 existing algorithms against VTAB, with all the models pre-trained on the ImageNet dataset (which isn’t included in VTAB). They evaluate on a range of image-based and patch-based models, as well as generative models like VAEs and GANs. Supervised learning models perform best with the highest performing model obtaining a mean score of 73.6% when using 1,000 training examples, and 91.4% when training on the full dataset. Though pre-training on ImageNet seems like a broadly good idea, it leads to fewer performance gains when models pre-trained on it are tested on specialized datasets, like ones from the medical domain, or on ones that require more structured understanding. 

Why this matters: Today, people are starting to train very large, expensive models on vast datasets in both the text and vision domains. Recently, these models have started to get very good, obtaining near-human performance on a variety of specific tasks. This has created the demand for sophisticated testing regimes to let us measure the capabilities of a given model on a diverse set of tasks so we can better assess the state of progress in the field. Such multi-task benchmarks have started to become common in various parts of NLP (typified currently by the GLUE and successor SuperGLUE) systems); VTAB is trying to do the same for text. If it becomes widely used, it will help us model progress in the vision domain, and give us a better sense of how smart our systems are becoming.
   Read more: The Visual Task Adaptation Benchmark (Arxiv).
   Get the code for the benchmark here (VTAB benchmark, GitHub).

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

Tech Tales: 

Hunting Big Game

How are we doing on the contract?, the boss asked. 

Bad, said the AI manager. The machines are fighting. 

What’s new?

The other side is winning. 

What? the boss asked. 

Neither of them said anything for a couple of seconds, and in the passing of those seconds both of them knew they had likely lost the negotiation: there was only one way to win in this game, and it was about having the biggest computer. If the other side was winning, then Global Corp – logically – must have a smaller computer. Unexpected? Yes. A situation they had taken taken every measure to avoid? Yes. But possible? Sadly, yes.
We lost the bid, said the AI manager. 

Where did it come from? The boss asked. 

I do not have a high-confidence answer here. 

I figured, said the boss. Maybe let’s work backwards – how many people could afford to outspend us?

There are around 50 organizations who could spend the same or greater than us, said the AI manager.

And other actors?

We estimate around 20 governments and perhaps 10 criminal groups might have the capacity also. 

Billionaires?

Separate from their corporations?

Yes. 

Perhaps 20. 

I leaned back in my chair and made a steeple with my fingers. 100 options. 100s of ways, both legal and illegal, to gather compute. A vast combinatorial space. I hoped I had enough computation to figure out who the person was before they acquired enough computation to put me at a permanent disadvantage in my business dealings.

   Assign all computers to searching for our likely adversary, I said. Buy computational puts as confidence increases and try and squeeze them out of the chip markets for a while. 

   You got it, said the AI manager. And somewhere on the planet, thousands of machines begin to whirr, trying to seek their counterparts. 

 

Things that inspired this story: High-frequency trading; Charles Stross’s ‘Accelerando‘; various research papers tying certain computational capabilities to certain scales or quantities of computation; the logical end-point of the ‘marketization’ of compute; the industrialization of AI; compute inequality; masters and servants and gods and monsters.