We are, of course, looking for ways to apply MuZero to real world problems, and there are some encouraging early results. To give a concrete example, internet traffic is dominated by video, and a big problem is how to compress these videos as efficiently as possible. You can think of this as reinforcement learning problem because there are these very complicated programs that compress video, but what you see next is unknown. But when you plug something like MuZero into it, our first results look very promising in terms of saving significant amounts of data, perhaps something like 5% of the bits used to compress a video.
In the longer term, where do you think reinforcement learning will have the greatest impact?
I am thinking of a system that can help you as a user achieve your goals most effectively. A really powerful system that sees all the things you see, that has all the same senses that you have, that is able to help you achieve your goals in your life. I think this is a really important question. Another transformation, in the long run, is something that could provide a personalized healthcare solution. There are privacy and ethical issues that need to be addressed, but this will have enormous transformative value; it will change the face of medicine and the quality of people’s lives.
Is there something that you think machines will learn to do in your lifetime?
I don’t want to put a timescale on it, but I would say anything a human can accomplish, ultimately I think a machine can. The brain is a process of calculation, I don’t think there is any magic in there.
Can we reach the point where we can understand and implement algorithms as efficient and powerful as the human brain? Well I don’t know what the schedule is. But I think the trip is exciting. And we should aim to achieve this goal. The first step on this journey is to try to figure out what it means to even achieve intelligence? What problem are we trying to solve by solving intelligence?
Beyond practical uses, are you convinced that you can move from mastering games like chess and Atari to real intelligence? What makes you think reinforcement learning will lead to machines with a common sense understanding?
There is a hypothesis, we call it the reward hypothesis, that’s enough, that says that the essential process of intelligence could be as simple as a system seeking to maximize its reward, and this process of ‘try to achieve a goal and try to maximize the reward. is enough to give birth to all the attributes of intelligence that we see in natural intelligence. It’s a hypothesis, we don’t know if it’s true, but it kind of gives direction to the research.
If we take common sense in particular, the reward sufficient hypothesis says well, if common sense is useful to a system, it means that it should actually help it to better achieve its goals.
Sounds like you think your area of expertise – reinforcement learning – is in some sense fundamental to understanding or “solving” intelligence. Is it correct?
I really see it as very essential. I think the big question is: is this true? Because it certainly goes against the way a lot of people view AI, which is that there is this incredibly complex collection of mechanisms involved in intelligence, and each of them them either has their own kind of problem to solve or their own way works, or maybe there isn’t even a clear definition of the problem for something like common sense. This theory says, no, in fact there can be this very clear and simple way of thinking about all intelligence, that is, it is a goal optimization system, and that if we let’s find a way to optimize the goals really, really well, then all these other things will come out of that process.