Max Welling, Research chair in Machine Learning at the University of Amsterdam and VP technologies Qualcomm Netherlands, talks about about machine learning, big data, artificial intelligence, historical data, deep learning and algorithms.
WHAT IS AI?
AI is a field in which we try to make machines more smart, and sort of traditional approaches to AI they code certain rules into systems and they let the machines reason over these rules. In more modern approaches to AI, which involves machine learning, we try to learn or make the machine more smart by learning it, having it learn from data or from historical data. You can view this perhaps as a system that has been programmed, the rules are fixed. If standing in front of a red light you know there could be a rule in a car that says if this traffic light is red, then you should break and stop.
But for machine learning you can view it as a program that constantly re-writes itself, so if it looks at data it will try to improve its code base, in order to become better and predicting the world after it's seen that data. It's much like humans learn, so they play around, they test hypotheses, they try things out and in that way they learn about the world and they make themselves smarter. Within that field of machine learning there is a sub field which is currently very hot, which is called deep learning. In deep learning we train a specific kind of model which is called a deep neural net, that is inspired by how the brain works. So let me explain a little bit how that works: in a brain we have neurons and we have connections between neurons, which are sort of information channels, that we call synopsis, and if the channel is sort of wide a lot of information can be transferred between two of those neurons.
When we learn, we change the capacity of these channels, so when these two neurons activate a lot together then we sort of make the channel bigger and more information can flow through it. In artificial neural network is exactly the same, so we have sort of layers of artificial neurons and in between we have connections and when the network sees data sets, tries to make predictions, for instance it looks at the image and it tries to see whether there's a car in the image. If it makes a mistake then there is a signal that says “you made a mistake” and then a signal propagates back through this neural network and tries to change the value of these channels to make them bigger or smaller, so that the next time around they look at the image they will actually make the correct prediction. Not all artificial intelligence is based on machine learning, on deep learning there's also fields that do not use learning to make machines smart, but I would say that trend is that most of AI now makes use of machine learning technology, because the best way to make machines smart is really by learning from data.
The traditional way to think of AI may be of robots that walk around in the world and interact with the world, but there's a lot of other ways that you can have smart machines or smart algorithms. For instance, on your phone there's algorithms that are guiding you through life, for instance it tells you where to go using your navigation system, but also recommender system so if you try to find your next movie on Netflix there’s a recommender that tells you what movies you might like or if you buy a book on Amazon and there's always recommenders. There's a lot of other ways in which algorithms are smart, for instance also search: if you search on Google that's a smart algorithm that not necessarily take the shape of a robot.
I think the most important thing that happened around 2010 is that we would get a huge computational power by the fact that we figured out that we could use GPUs which are processors doing the video rendering in computers, and we could leverage those to train large neural networks. It's particular to compute power which has been driving the progress in artificial intelligence, at the same time of course a lot more data became available as well, and these models only start performing very well when there's a lot of data available.
All of that fueled into much bigger and stronger models, with lots more parameters that we could train. For instance, we can now train models with almost a thousand layers, where before we were stuck with maybe a few layers, and so the extra flexibility, any extra power that we have in these neural nets make the models a lot stronger.
Artificial intelligence, on one hand you consider it as an efficient computational system, there’s certainly algorithms that are not more than that. You can think of simple search methods or as simple navigation methods for your car, GPS methods, but there's also algorithms which are starting to behave a lot more like real intelligence to the way we would consider real intelligence, like Alfa Go, which is the engine that beat the world champion in the game of Go. There we saw quite sophisticated moves, in fact they looked highly original and they would maybe be considered even superior to human Go players.
I think when the systems become more and more complex, we will see intelligent behavior that is recognized as either human or superior to human. You could also imagine that they look a bit alien to us, so maybe it's a different kind of intelligence than humans have and so it's not necessarily that is if the system becomes very intelligent, that it is an intelligence that we can easily relate with.
AI currently is very good in doing predictions, that are sometimes better than humans but in a very limited domain. There are systems now that can diagnose skin cancer better than any dermatologist can, similar things have happened in digital pathology and other sort of limited domains. If you have a lot of data, restrict yourself to a limited domain than we can build systems that are extremely good. Also Alfa Go is an example because there because its domain is very limited, but they're not very good at flexible behavior like humans, they can behave intelligently in many different kinds of environments, even in environments that they have never ever seen before, so the current AI systems don't have that type of flexibility. In other words, they don't understand enough about the world yet in order to behave intelligently in different environments.