Import AI: Issue 4: Medical-grade machine learning in Uganda, free data, and a personal announcement.

by Jack Clark

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Intelligence and compression: being able to summarize something is a key trait of intelligence, so new work from Google that shows how to use neural networks for image compression is worth paying attention to. The paper, ‘full resolution image compression with recurrent neural networks’, outlines ways to compress images using neural networks, and appears to be competitive with or outperform existing techniques. The difference is that the neural networks have been taught to compress things through understanding what they are compressing, rather than programmed with specific knowledge of how compression works.

Please build this for me: Neural networks can do compression.  They can alsoscale up low-resolution images, given a sufficiently broad dataset. Twitter’s recent acquisition Magic Pony had developed some good technology in this area. Now I find myself wondering why there isn’t a web service that can take my pictures and scale them up for me for a (small) fee? This would be handy for landscape and/or tourist shots where there’s already a lot of data out there. I suspect Google will eventually add this feature to Google Photos. [Post-publication edit: It turns out this does already exist –]

Free data! Data is the crude oil of AI. Just as the world relies on oil the development of AI relies on access to data. If you don’t have data, you don’t have the raw material needed to develop, experiment on, and enhance AIsystems. So Kaggle should be congratulated for creating ‘Kaggle Datasets’, which hosts (free!) data, and lets people upload their own.

Free tools! Facebook has released fastText, some open source software for text and word classification, making it easier for people to build software that analyzes the sentiment of a piece of text, or figure out how a previously unseen word relates to known words. It’s hard to think of another industry which makes so many of its tools available so freely.

Better tools!: plumbing is important. You can think of a neural network as an intricate stack of many layers, each containing many vessels, connected to one another by pipes. Liquid flows from the bottom layer of the system to the top, then washes back down again, altering numbers associated with each vessel along the way. Once it reaches the bottom, the process starts all over again. What if you didn’t need to wait for it to wash back down? That’s the essence of a new paper from Google DeepMind called ‘decoupled neural interfaces using synthetic gradients’. It outlines a way to train very large neural networks more rapidly by being able to unhook some of the computations from one another. This is useful because it lets you do more experiments in the same amount of time, speeding progress. Some wonder if it will mostly work at Google-scale (translation: big. Like, Salvador Dali mind-bendingly weird Big). and will not be so useful for smaller systems. “The obviousness of it makes me think it is something others who have worked long and hard in the field have thought of but never had the resources to execute. But then, some things are obvious only in retrospect.” writes developer Delip Rao.

Don’t believe what you see: new computer vision techniques are making it much, much easier for people to manipulate images and videos. This will lead to new forms of propaganda, art and, of course, memes. Earlier this year researchers demonstrated ‘Face2Face’, which lets you manipulate the faces of people by mapping your expressions onto theirs, so you can literally put words into someone else’s mouth. Now, a video from a MIT PHD student named Abe Davis shows us ‘Interactive Dynamic Video’, technology to manipulate and animate objects from video. All you need is a slight vibration. The system works by looking at how vibrations propagate through an object and then use that to figure out the underlying structure of it. This could give people a way to analyze the structural integrity of bridges by filming them in a stiff breeze, or make it easy add CGI effects to films by making certain objects interact with each other.

An all-seeing, all-thinking, globe-spanning eye: satellite company Planet Labs has partnered with the computer vision experts at Orbital Insight to go after customers in the financial sector, pairing a (growing) fleet of around 60 satellites with a team teaching computers how to read the (literal) tea leaves. That will make it easier for investors to ask questions like ‘what are the shipping trends, based on the traffic at Major Hub X’, or ‘how will the biofuel market respond to crop fluctuations in Country Y’. As usage of this type of technology grows the price should come down, letting people like you and me analyze our own local environment. I’m particularly interested in exploring how the colors of the vegetation on the East Bay hills respond to seasonal temperature and rainfall fluctuations.

The doctor will (really) see you now: AI is going to make healthcare much, much better. The same computer vision algorithms that companies like Google have been busily refining for years to correctly classify (and serve ads against) images, are also perfectly suited to analyzing and labeling medical imagery.Stanford researchers have trained an AI system to figure out if something was cancerous or not. Its system learns to categorize nearly 10,000 individual traits compared to the several hundred a (human) pathologist might use. “These characteristics included not just cell size and shape, but also the shape and texture of the cells’ nuclei and the spatial relations among neighbouring tumor cells”. Meanwhile, Ugandan researchers have used similar techniques to attain good performance at spotting intestinal parasite eggs in stool samples, diagnosing malaria in thick blood samples, and tuberculosis in sputum samples. “The fact that in our experiments the same network architecture successfully identifies objects in three different types of sample further indicates its flexibility; better results still are likely given task-specific tuning of model parameters with cross validation for each case,” they write.

OpenAI imports Jack: I’ve joined OpenAI. I’ll be starting in a few weeks as our ‘Strategy and Communications Director’, which basically entails explaining AI to the world, whether that be journalists, researchers, regulators, or other interested parties. Suggestions? Questions? I shall continue to write this newsletter in my spare time.

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at or tweet at me@jackclarksf