Deriving the ELBO: The Loss Function Behind Variational Autoencoders
If you've trained a VAE, you've used the ELBO whether you noticed it or not. Here I derive it from scratch, starting from a plain autoencoder.
Blog
I like to write about the things I do and share them :) RSS
Guaranteeing Your Model's Error Rate with Conformal Prediction
Your model says "cat", but is it 99% sure or 1% sure? Conformal prediction wraps any trained model with a coverage guarantee, without retraining. Intuition, proof and a from-scratch implementation.
Visualizing Distribution Change in Forward Diffusion Process
A simple experiment showing how an input's distribution changes as Gaussian noise is added step by step, following the DDPM paper.
Building a Tic Tac Toe Agent with Proximal Policy Optimization
Building the environment, the policy and the PPO training loop for a tic-tac-toe agent, one piece at a time.
Residual Connections & Gradient Flow
Do residual connections actually help deep networks train? I measured the gradient flow myself.
Simple Positional Encoding with PyTorch
Why transformers need positional encoding, how it works, and a minimal version written from scratch in PyTorch.