I am largely in debt of quality documentation, regular project updates, a project website and so on. I mostly prioritize development work, because I want things to be better faster. But… Here goes one well overdue project update on Polars. Quick reminder Polars is an in-memory DataFrame library written in Rust and exposing an API in Rust and Python. It has tightly interopability with Apache Arrow and uses arrow memory as columnar store. Read more

Update 1 Aug 2021 Benchmarks age poorly, a few months of work later Polars is the fastest performing solution in the db-benchmark. Join performance 14x pandas 2x R data.table Groupby performance 6-20x pandas 1-2x R data.table db-benchmark results 1. Introduction At the time of writing this, the coronavirus has been in our country for a year, which means I have been sitting at home for a very long time. Read more

This is post was a real eye-opener for me with regard to the methods we can use to train neural networks. A colleague pointed me to the SLIDE[1] paper. Chen & et al. discussed outperforming a Tesla V100 GPU with a 44 core CPU, by a factor of 3.5, when training large neural networks with millions of parameters. Training any neural network requires many, many, many tensor operations, mostly in the form of matrix multiplications. Read more

This post we will explore a type of normalizing flow called **Inverse Autoregressive Flow**. A composition (flow) of transformations, while preserving the constraints of a probability distribution (normalizing), can help us obtain highly correlated variational distributions. Don’t repeat yourself If what was mentioned in the previous lines didn’t ring a bell, do first read these posts: variational inference and normalizing flows. This post could really be seen as an extension of the latter. Read more

Four of my last five blog posts were more or less related to Baysian inference with variational methods. I had some momentum, and I wanted to use the traction I gained to do another post (which will come!) on enhancing variational methods with Inverse Autoregressive Flows (IAF), but first I have to get something different out of the way. In the paper describing IAF, they refer to an autoregressive neural network (and further assume his to be clear knowlegde). Read more

Last posts we’ve investigated Bayesian inference through variational inference (post 1/post 2). In Bayesian inference, we often define models with some unknown model parameters $Z$, or latent stochastic variables $Z$. Given this model and some observed data points $D = \{ D_1, D_2, \dots, D_n \} $, we are interested in the true posterior distribution $P(Z|D)$. This posterior is often intractable and the general idea was to forgo the quest of obtaining the true posterior, but to accept that we are bounded to some easily parameterizable approximate posteriors $^*Q(z)$, which we called variational distributions. Read more

In the posts Expectation Maximization and Bayesian inference; How we are able to chase the Posterior, we laid the mathematical foundation of variational inference. This post we will continue on that foundation and implement variational inference in Pytorch. If you are not familiar with the basis, I’d recommend reading these posts to get you up to speed. This post we’ll model a probablistic layer as output layer of a neural network. Read more

Not that long ago I wrote an introduction post on Gaussian Processes (GP’s), a regression technique where we condition a Gaussian prior distribution over functions on observed data. GP’s can model any function that is possible within a given prior distribution. And we don’t get a function $f$, we get a whole posterior distribution of functions $P(f|X)$. This of course, sounds very cool and all, but there is no free lunch. Read more

Bayesian modeling! Every introduction on that topic starts with a quick conclusion that finding the posterior distribution often is computationally intractable. Last post I looked at Expectation Maximization, which is a solution of this computational intractability for a set of models. However, for most models, it isn’t. This post I will take a formal definition of the problem (As I’ve skipped that in the Expectation Maximization post) and we’ll look at two solutions that help us tackle this problem; Markov Chain Monte Carlo and Variational Inference. Read more

I wanted to learn something about variational inference, a technique used to approximate the posterior distribution in Bayesian modeling. However, during my research, I bounced on quite some mathematics that led me to another optimization technique called Expectation Maximization. I believe the theory behind this algorithm is a stepping stone to the mathematics behind variational inference. So we tackle the problems one problem at a time! The first part of this post will focus on Gaussian Mixture Models, as expectation maximization is the standard optimization algorithm for these models. Read more