Blog
- (2024-10-07) paper thoughts -- questionable research practices (QRPs) in machine learning
-
(2024-05-31) EDM diffusion models - a Jupyter implementation, and how they are implemented in practice
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.
-
(2023-04-27) A deep dive into conditional variational autoencoders
An excerpt from my PhD thesis. We discuss various difficulties involved in correctly training variational autoencoders, as well as elucidate their behaviour from a theoretical perspective.
-
(2023-04-15) Vicinal distributions as a statistical view on data augmentation
An excerpt from my PhD thesis. I introduce data augmentation through the lens of vicinal distributions, and introduce mixup augmentation as a multi-sample generalisation of it.
- (2023-03-20) My notes and derivations for SMLDs
-
(2023-01-27) Techniques for label conditioning in Gaussian denoising diffusion models
A discussion of three recently-proposed conditional variants for Gaussian diffusion probabilistic models -- classifier-based guidance, classifier-free guidance, and the conditional ELBO (evidence lower bound).
-
(2022-09-24) Learning the conditional prior over classes for image diffusion
My implementation of a conditional diffusion model for speech enhancement, demonstrated for images on the MNIST dataset. Unlike other conditional formulations which use classifier-style guidance, this proposes a conditional variant of the ELBO.
- (2022-08-02) The obsession with SOTA needs to stop
- (2022-07-24) Towards a more sane Mac OS user experience, and I am late to the party
-
(2022-07-11) My notes on discrete denoising diffusion models (D3PMs)
My notes on D3PMs. It includes derivations for the two main equations presented in their work.
- (2021-08-17) A somewhat mathematical introduction to Gaussian and Laplace autoencoders
- (2021-06-28) Training GANs the right way