One of the things that I find is usually missing from many ML papers is how they relate to the fundamentals. There's always a throwaway line where it assumes something that is not at all obvious (see my post on Importance Sampling). I'm the kind of person who likes to understand things to a satisfactory degree (it's literally in the subtitle of the blog) so I couldn't help myself investigating a minor idea I read about in a paper.

This post investigates how to use continuous density outputs (e.g. a logistic or normal distribution) to model discrete image data (e.g. 8-bit RGB values). It seems like it might be something obvious such as setting the loss as the average log-likelihood of the continuous density and that's almost the whole story. But leaving it at that skips over so many (interesting) and non-obvious things that you would never know if you didn't bother to look. I'm a curious fellow so come with me and let's take a look!

Read more…

It's been a long time coming but I'm finally getting this post out! I read this paper a couple of years ago and wanted to really understand it because it was state of the art at the time (still pretty close even now). As usual though, once I started down the variational autoencoder line of posts, there was always yet another VAE paper to look into so I never got around to looking at this one.

This post is all about a proper probabilistic generative model called Pixel Convolutional Neural Networks or PixelCNN. It was originally proposed as a side contribution of Pixel Recurrent Neural Networks in [1] and later expanded upon in [2,3] (and I'm sure many other papers). The real cool thing about it is that it's (a) probabilistic, and (b) autoregressive. It's still counter-intuitive to me that you can generate images one pixel at at time, but I'm jumping ahead of myself here. We'll go over some background material, the method, and my painstaking attempts at an implementation (and what I learned from it). Let's get started!

Read more…

It took a while but I'm back! This post is kind of a digression (which seems to happen a lot) along my journey of learning more about probabilistic generative models. There's so much in ML that you can't help learning a lot of random things along the way. That's why it's interesting, right?

Today's topic is importance sampling. It's a really old idea that you may have learned in a statistics class (I didn't) but somehow is useful in deep learning, what's old is new right? How this is relevant to the discussion is that when we have a large latent variable model (e.g. a variational autoencoder), we want to be able to efficiently estimate the marginal likelihood given data. The marginal likelihood is kind of taken for granted in the experiments of some VAE papers when comparing different models. I was curious how it was actually computed and it took me down this rabbit hole. Turns out it's actually pretty interesting! As usual, I'll have a mix of background material, examples, math and code to build some intuition around this topic. Enjoy!

Read more…

This post is going to be about a really simple idea that is surprisingly effective from a paper by Bagherinezhad et al. called Label Refinery: Improving ImageNet Classification through Label Progression. The title pretty much says it all but I'll also discuss some intuition and show some experiments on the CIFAR10 and SVHN datasets. The idea is both simple and surprising, my favourite kind of idea! Let's take a look.

Read more…

"In theory, theory and practice are the same. In practice, they are not."

I read a very interesting paper titled ResNet with one-neuron hidden layers is a Universal Approximator by Lin and Jegelka [1]. The paper describes a simplified Residual Network as a universal approximator, giving some theoretical backing to the wildly successful ResNet architecture. In this post, I'm going to talk about this paper and a few of the related universal approximation theorems for neural networks. Instead of going through all the theoretical stuff, I'm simply going introduce some theorems and play around with some toy datasets to see if we can get close to the theoretical limits.

(You might also want to checkout my previous post where I played around with ResNets: Residual Networks)

Read more…