PhD candidate at MILA and Polytechnique Montreal. Interested in self-supervised learning, generative models, few-shot learning, and inverse graphics.
I’ve been thinking about offline model-based optimisation for the past six months, mostly in the background while working full-time, trying to write up something that captures both what I found and what still needs to be done in the field. It’s about validation, extrapolation, brittle surrogates, and how we evaluate generative models when the ground truth is expensive.
my thoughts on the paper "questionable research practices in machine learning" by leech et al
I wrote a self-contained implementation of NVIDIA's EDM diffusion model in a Jupyter notebook, as well as its associated sampling algorithms. I also discuss the rather confusing names used for real-world implementations of those algorithms.
VAEs can be difficult to understand and get working right. In this post I will dive into the theory of them, derive interesting equations which elucidate their behaviour, corroborate those insights on a simple toy dataset, as well as discuss how their issues can be alleviated with adversarial methods.
Vicinal distributions as a statistical view on data augmentation
notes, derivations, and thoughts I have written to help me understand score-based generative models via Langevin dynamics (SMLDs).
Techniques for label conditioning in Gaussian denoising diffusion models
Learning the conditional prior over classes for image diffusion
The obsession with SOTA needs to stop
My journey in trying to craft a better user experience for myself on Mac OS X.