Import AI: Issue 6: Amazon’s New UK AI Team, Baidu’s Frameworks, and an OpenAI Member’s Q&A
by Jack Clark
Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.
Baidu Paddles into AI Software: ‘It is a truth universally acknowledged, that an ambitious company in possession of a large technology team must be in want of its own artificial intelligence software’ – Jane Austin. No surprise then that Chinese technology giant Baidu announced its own free deep learning framework named ‘Paddle’, last week. Deep learning frameworks make it easier to design and build AI software, saving programmers time. Paddle will compete with similar software from Google (TensorFlow), Microsoft (CNTK), Amazon (DSSTNE), academic projects like Torch and Theano, and a smorgasbord of other add-ons, libraries, and tools from other companies. These tools have becoming marketing devices in their own right; Baidu announced Paddle at Baidu World in Beijing last week, ahead of the full release of the documentation, which is due for Sep 30th. Now taking bets on when Apple will release its own open source AI software…
There Can Only Be Three or Four: AI software will likely go the way of Linux distributions, with a few popular ones garnering the vast majority of users, and a long, long tail of less widely used ones extending out as far as the eye can see. Companies are likely going to compete heavily to gain the top spot(s), as that will make it easier for them to sell compatible cloud services to developers and to identify new talent to hire. This Hacker News poll shows Google’s TensorFlow and Francois Chollet’s Keras to be in the lead for now.
Amazon Builds UK AI Team: Amazon is building a new machine learning group in the UK and has persuaded Neil Lawrence of the University of Sheffield to join it. Multiple large companies have previously approached Mr Lawrence about joining, so Amazon has pulled off something of a coup here. The group will work alongside Ralf Herbrich’s group in Berlin. We hope Neil Lawrence will continue to write about AI for the research community and general public. He says he plans to. All the best!
Quantum AI Could Be Closer Than You Think: Experts don’t know if truly intelligent machines will require quantum computers, but a quantum computer certainly wouldn’t hurt, because the machines can solve problems that today’s computers simply can’t handle. Companies like Microsoft, Google, and IBM, are all looking into using quantum computers for applications in security and simulation, as well as to improve computer vision techniques and more. Now Google plans to use a quantum computer to simulate the behaviour of a random arrangement of quantum circuits. If it manages to do that, then the company will have made a significant step in tackling tasks that exceed the capabilities of traditional computers, say experts interviewed by New Scientist.
Industrial-scale Recommendations: Google has published a paper outlining the recommender system that underpins YouTube. It does some interesting things with regards to the age of videos to help it cotton-on to popular trends.
The (Real) Future of AI: Stanford has released the 2016 report of its One Hundred Year Study on Artificial Intelligence. Experts like Rodney Brooks (Rethink Robotics), Oren Etzioni (Allen Institute for AI), Astro Teller (Google X), Eric Horvitz (Microsoft), and many others have analyzed AI and tried to make some informed projections about where it is going. The answer includes autonomous cars, smarter home robots, greater health care systems, and the usual fuzzy projections about how it may influence employment. “AI will gradually invade almost all employment sectors, requiring a shift away from human labor that computers are able to take over,” they write. From a policy level it’d be great to have more funding for interdisciplinary study of the societal impacts ofAI, as well as clarifying how some rules (like the DMCA) could influence AIresearch, and employing more AI experts within government, they write. Download the report here.
Oooh, Fashion!: A website from Google and Zalando uses machine learning, massive amounts of data mining, and the input of over 600 fashion ‘professionals and influencers’, to create a tool that can procedurally generate new fashion items. Project Muze’s results are rather underwhelming, but it’s notable that both parties appeared to invest so many resources into the project. More to come, I expect.
You Have Questions, They Have Answers: Andrej Karpathy of OpenAI is doing a Quora Session on Thursday. Ask away.
Deep Learning Education: Jeremy Howard of Kaggle/Enlitic/Fast.AI has launched a Deep Learning course with the University of San Francisco. “We think this is the first ever in-person university-accredited deep learning certificate course. Only real prerequisite is decent coding expertise, although some memory of high-school level linear algebra (matrix-matrix products) and calculus (the chain rule) would be helpful,” he says. “There are scholarships available for anyone who isn’t able to pay for the course – although we’ve tried to keep the cost down too. (And I’m donating my teaching fees.)” Check out the course (and a good introductory deep learning video) at this link.