My name is Nicolas. I am a researcher at Leapmind.
As a researcher it is my job to stay up to date with the newest developments in the field of deep learning, put real real life problems in the corresponding technological context and work together with the engineers on the implementation of deep learning algorithms.
Today I would like to talk a little bit about the development in this field that received the most attention this year.
I am of course talking about DeepMinds AlphaGo.
As a hobby player of Go myself, this is also one of the things that motivated me to start a career in Deep Learning.
Here at Leapmind we work extensively on applying deep learning algorithms to image processing. Although this seems to be very unrelated to the problem of creating an AI to play Go, both rely heavily on so called Deep Convolutional Neural Networks.
Convolutional Neural Networks have been dominating the progress in image processing over the past few years.
They work by taking a large input (for example more than 3 million parameters in the case of a 1024x1024 pixel color image) and create an abstract representation of the image through millions of operations that can be used to simplify tasks such as classification or segmentation. Machine learning is used for the interpretation of these representations as well as in order to create them by learning structures from a large amount of samples.
Now, how is this related to the problem of creating an AI to play Go. Go is an extremely complex game which is often being emphasized by the fact that there are more possible board positions than there are atoms in the universe.
In AlphaGo Convolutional Neural Networks came in by interpreting a Go position as a 19x19 image (size of a go board) to create abstract interpretations mentioned above, and hence drastically decreasing the complexity of the problem, along with applying other advanced state of the art techniques such as reinforcement learning.
The success of AlphaGo seems somewhat frightening when thinking in science fiction terms of a world were machines will have surpassed humans. People with some experience in playing Go will be particularly intrigued to feel in this way since Go players due to the complexity of the game rely heavily on intuition (unlike in Chess where the study of specific reoccurring board positions is significantly more important), and Go has been promoted for centuries as a game from which, by applying its strategies and tactics to other problems, we can learn a lot for our lives.
It is therefore very important to keep in mind that AlphaGo is fundamentally unable to teach itself new strategies for learning and I hope that it comes without saying to the reader that it won’t suddenly take control of our lives by, for example, applying its skills to trading on the stock market.
I also want to mention that, although Convolutional Neural networks are inspired by the human brain, the representations of structures that they learn are very different from the way we humans abstract certain patterns. Also, a first AI system called Differentiable Neural Computer that learns how to reason based on memory was the next breakthrough from DeepMind later this year. It is, for example able, to find the quickest way from one station to another based on pictures of the London underground system. But this sounds a lot less frightening now, doesn’t it?
Even though recent results can be breathtaking, we are still at the very beginning of the road that might eventually lead to creating true AI, and that’s why right now is such an amazing time to work in Deep Learning! 🙂