Károly Zsolnai-Fehér explains how to visualize what neural networks learn, from simple edges and colors in early layers to complex patterns in deeper ones. Christopher Olah's Distill article collects the latest visualization techniques with interactive web demos you can try yourself.
Dear fellow scholars, this is two minute papers with this one is going to be a treat. As you know all too well after watching at least a few episodes of this series, neural networks offer us amazingly powerful tools to defeat problems that we didn't stand a chance against for a long, long time. We are now in the golden age of AI and no business or field of science is going to remain unaffected by this revolution. However, this approach comes with its own disadvantage. Compared to previous handcrafted algorithms, it is harder to know what is really happening under the hood. That's also kind of the advantage of neural networks because they can deal with complexities that we humans are not built to comprehend. But still, it is always nice to peek within a neural network and see if it is trying to learn the correct concepts that are relevant to our application. Maybe later we'll be able to look into a neural network, learn what it is trying to do, simplify it, and create a more reliable handcrafted algorithm that mimics it. What's even more, maybe they will be able to write this piece of code by themselves. So clearly there is lots of value to be had from the visualizations. However, this topic is way more complex than one would think at first. Earlier we talked about a technique that we called activation maximization, which which was about trying to find an input that makes a given neuron as excited as possible. Here you can see what several individual neurons have learned when I trained them to recognize wooden patterns. In this first layer it is looking for colors. Then in the second layer some basic patterns emerge. As we look into the third layer we see that it starts to recognize horizontal, vertical and diagonal patterns. And in the fourth and fifth layers it uses combinations of the previously seen features and as you can see, beautiful somewhat symmetric figures emerge. If you would like to see more on this, I put a link to a previous episode in the video description. Then a follow up work came for multifaceted neuron visualizations that unveiled even more beautiful and relevant visualizations. A good example was showing which neuron is responsible for recognizing groceries. A new distilled article on this topic has recently appeared by Christopher Ola and his colleagues at Google. This tale is a journal that is about publishing clear explanations to common interesting phenomena in machine learning research. All their articles so far are beyond amazing, so make sure to have a look at this new journal as a whole. As always, the link is available in the video description. They usually include some web demos that you can also play with. I'll show you one in a moment. This article gives a nice rundown of recent works in optimization based feature visualization the optimization part can take place in a number of different ways, but it generally means that we start out with a noisy image and look to change this image to maximize the activation of a particular neuron. This means that we slowly morph this piece of noise into an image that provides us information on what the network has learned. It is indeed a powerful way to perform visualization, often more informative than just choosing the most exciting images for a neuron from the training database. It unveils exactly the information the neuron is looking for, not something that only correlates with that information. There is more about not only visualizing the neurons in isolation, but getting a more detailed understanding of the interactions between these neurons. After all, a neural network produces an output as a combination of these neuron activations, so we might as well try to get a detailed look at how they interact. Different regularization techniques to guide the visualization process towards more informative results are also discussed. You can also play with some of these web demos. For instance, this one shows the neuron activations with respect to the learning rates. There is so much more in the article. I urge you to read the whole thing. It doesn't take that long and it is a wondrous adventure into the imagination of neural networks. How cool is that? If you have enjoyed this episode, you can pick up some really cool perks on Patreon, like early access, voting on the order of the next few episodes, or getting your name in the video description as a key contributor. This also helps us make better videos in the future, and we also use part of these funds to empower research projects and conferences. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time. Ra.
