Presentation: Understanding Deep Learning

I frequently use deep learning for a range of different things, very often with time series. Previously I worked in finance. We tried to predict different kinds of stock prices, bond prices and economic indicators. I also worked with self-driving cars. The inspiration for this talk was really very much based in what I was doing because I found that when I first started working deep learning as opposed to other kind of machine learning models, I had this spreadsheet and then I would try to find the best model, for example, for predicting stock prices. And then I would try this one model architecture and then I would try hundreds of configurations of different numbers of layers, numbers of units in each layer, different kind of parameters, different kinds of optimizers. Then I felt that at the time I was just doing trial and error and I spent a lot of time running a lot of experiments without a clear sense of the direction in which I was going. I didn't know how the different parameters of the different components affected the output, how good the model was. When I did understand it many months later I thought, if I'd only known this then I could have saved so much time running those hundreds of experiments, each of which took hours. So I thought it would be a really good thing to talk about.

We're going to start from the really fundamental stuff, but we're going to talk about it in a way that I hope people who have been using neural networks for months can still find it helpful. Neural networks have a lot of layers, so we're going to start off by talking about what happens in a single layer - a linear layer and a non-linearity - and we're going to talk about what kinds of things one or a few layers can model, why you need a non-linearity, and what different non-linearities do. And then we're going to move onto models with more and more layers because the consensus seems to be that the deeper the models are the better. But often it's very difficult to train very deep models. I'm going to go through the usual problems, when those problems might arise, and how people have been tackling those problems in state of the art models. Then I'll finish with practical tips to train really good models.

One of the things I'm going to talk about is about one parameter that's really important that's called the learning rate. If you don't set the learning rate right, if you set it to be too high then your model is going to jump from one solution to another. It will be very unstable and you really don't want that. Or maybe it's too low, then your model is not going to learn anything. And some other practical things. In the ML community people have found there's a trick called batchnorm that improves the performance of the model a lot. On the day I'll talk about more quick tips that the research community has figured out through years of experimentation.

I wrote my talk to be informative enough for the person who has been working for deep learning for a while but hasn't really understood it, but I really want this talk to be understandable to someone who has not done any deep learning at all. So it will be simple enough for people who have not had any exposure to DL to understand. But also I hope it'll be insightful enough such that even if you've done machine learning for a while and don't have a huge understanding of deep learning that it will still be useful to you.

I want them to leave the talk with an understanding of deep learning, but then the question is what does understanding of deep learning mean, right? In terms of the takeaways, my hope is that firstly they can understand the building blocks of neural networks, what they're made of, how they're trained. And secondly that they can understand why some hacks improve performance.

The overarching theme of it is the hope that understanding these two things can give them the intuition on how to build a model, how to improve it, how to design and to debug these models, because as with a lot of software engineering the most annoying thing and the thing that takes a lot of time is debugging these models, and it makes a big difference if you can use your intuition to figure out what might be going wrong. So I think the main takeaways would be having the confidence to do that, and being able to build on this foundation, being able to pick up new architectures quickly. Those would be the main takeaways.