As part of our ongoing AI In your Community series, I talked to François Chollet – author of the open source neural network framework Keras, co-founder of AI-ON (Artificial Intelligence Open Network), creator of a learning platform for Artists – Wysp, working on Deep Learning, Google.
François Chollet
TC: How can children learn more about AI?
FC: One fun way to get started is to build very simple video games in JavaScript or modify an existing game. There’s a very low barrier to entry and it can be a gateway to learning about AI, because once you have a video game, you can try to write a basic A.I. that plays the game.
TC: That is intriguing and something we have actually not thought of! Thank you! So why did you start AI-ON?
FC: AI-ON is a collaborative, open research project. I started it with two other colleagues. The goal is to organize teams of people (who may not know each other) to do academic research on an A.I. research problem. We want to produce significant scientific contributions in a distributed and open source fashion — and this idea is a new one. Until now no one has really applied the open source philosophy and process to science. We know it works very well for software development and in many ways the scientific process is easier and simpler than software development, so I believe it should be successful.
TC: Why do you say that?
FC: Well, software development is extremely challenging. For instance, in A.I. and machine learning, most papers are written by small groups of researchers, maybe just two or three . But significant software projects cannot be built by just one or two people and the standards of success for software are more difficult to meet than the standards of success for a scientific paper. Having worked on both writing scientific papers and building software projects, I find that building software projects scares me!
We think that the open source process can be applied to producing scientific knowledge, especially in machine learning and AI. Also, we want to give people, who would not normally have an opportunity to do the work of a scientist, access to the scientific process — to the experience of working with other researchers on an important project, running experiments, designing solutions to problems, and finally publishing a paper. Currently, the only way you get to work on such projects is to join a graduate school. There are many people who would be qualified to do scientific research but they either could not or do not want to go to graduate school. So we are hoping that with AI-ON, we can provide them an opportunity to advance AI research.
A similar model has been successfully applied by Kaggle — a website through which companies organize machine learning competitions and even non-academics outperform state of the art approaches published by academics. The typical profile of such a competitor is a software engineer working at a small company, without a lot of machine learning experience. But they are very creative, driven and able to learn stuff on the fly to solve the problems that they’re trying to solve. We’re trying to harness the talent and drive of such people to advance scientific progress in a meaningful way. The open source process could be a way to harness this talent, to get these people into science, and also to start advancing science in a distributed, decentralized way outside academia.
TC: So what inspires you and drives your interest?
FC: My interest in AI is a bit philosophical. I don’t see it as a way to get rich, or a way to do something big, as they say “change the world” or “make the world a better place.” It’s not that. For me, it’s mostly a way to understand who we are, what we are — it’s AI as a way to understand intelligence and the human mind. That’s something I care about deeply and is real for me. I will never get bored doing it because it is about trying to pierce the veil, understand the nature of reality, the nature of who we are. And of course if you are able to solve the problem of intelligence, then you are able to build what’s next, to build what comes after us.
TC: How long has AI-ON been running?
FC: It’s been one year and a few months. We’ve had a few nice successes. We’ve seen people successfully build teams and projects. The missing piece is that we haven’t published any new papers.
TC: Have you collected any data on what makes a successful team?
FC: Yes. What we learned in the first year was that it is critical to have a project owner for each project. The project owner is someone who really cares about the project, is the CEO, and is always going to be there to onboard new participants, to guide people in terms of questions, and so on. It turns out that for an open research project to be successful, it cannot be self-organized. The project has to be structured in the classical, hierarchical way. It can be spatially distributed but it needs to be fairly centralized in terms of relationships.
TC: That’s very interesting! How is it that some open-source software projects are successful? I thought they were self-organized.
FC: Actually not. All successful, open source software projects are extremely centralized and hierarchically structured. In fact, the most successful ones tend to be backed by companies. So in reality, AI-ON in its first year failed because of that. Our project did not have strong ownership. People just cannot come together and simultaneously start collaborating in a meaningful way. They need to be guided from above. They need a manager who has a global view of the project, who can answer questions, who can tell people what to do. If you don’t have that, your project is not going to succeed.
TC: Interesting. So what role does AI-ON play?
FC: Our role is to mentor project owners, to tell them how they can make their project succeed, based on what we learned. We also connect participants with project owners and serve as the intermediary.
TC: Looking into the future, what field of AI excites you?
FC: I’m very excited about Program Synthesis, which is the automation of software development — figuring out how to automatically write programs that solve specific problems. This goes far beyond the sort of machine learning that we’ve been doing so far, and that is an area of AI that’s growing very fast and gaining more and more interest.
TC: So what is the current state of Program Synthesis?
FC: It’s very much ground zero. We know nothing; everything needs to be built. And that’s what makes it so exciting!
