Latest Posts

    Toward a Learning Theory Perspective on the Evolutionary Benefit of Sexual Reproduction

    April 29, 2024

    The benefit of sexual reproduction is a long-standing question of evolutionary biology. Surveying the literature gives the impression that the question remains far from resolved. The closest thing I found to a consensus view is something along the lines of

    Sex facilitates the accumulation in a single individual of advantageous mutations which arose separately in different individuals.

    Why variational autoencoders?

    December 21, 2023

    Variational inference, and in particular its application to variational autoencoders (VAEs; Kingma & Welling, 2013; Rezende et al. 2014), is a powerful tool for learning to approximate complex probability distributions from data. However, the way it is often introduced initially made it seem somewhat mystical to me. In this post, I will try to introduce variational inference in a way that shows how it might arise from first principles as a natural solution to learning a neural-network-based latent-variable model. I will also highlight some of the nuances and subtleties that are often glossed over and motivate why we might want to learn a latent-variable model in the first place compared to other alternatives. This post will be a bit technical and assumes some familiarity with probability theory and machine learning.