Back to Projects

Deconstructing Autoencoders from Scratch

Explore the mathematics of information bottlenecks, compression, and representation learning — built entirely from scratch in pure PyTorch.

Part 1

The Math of Compression

Breaking down the information bottleneck, reconstruction loss, manifold hypothesis, and the relationship between undercomplete and overcomplete representations.

Part 2

PyTorch Implementation

Building four variants — Vanilla, Denoising, Sparse, and Convolutional Autoencoders — entirely from scratch.
View Code on GitHub

Part 3

What the Bottleneck Learns

Visualizing 2D latent spaces, denoising quality, sparsity patterns, and reconstruction fidelity across all four variants.