mnist-from-scratch

Overview Link to heading

A high-performance, educational machine learning laboratory for the MNIST dataset, built entirely from scratch in Rust. This project implements fundamental ML architectures without high-level frameworks, focusing on raw mathematical implementation and terminal-based visualization.

Ambition Link to heading

Work through neural-network fundamentals from first principles in Rust, with the code structured so the math and learning dynamics stay legible.

What’s novel Link to heading

  • Implements MNIST training as a fundamentals-first exercise rather than wrapping a large ML framework.
  • Uses Rust to make the learning pipeline explicit and inspectable.
  • Positions the repository as both a learning artifact and a base for deeper ML experiments.

Highlights Link to heading

  • Custom Dataset Pipeline: Hand-written parser for the IDX binary format.
  • Perceptron: Multiclass mistake-driven linear classifier.
  • Softmax Regression: Probabilistic linear model with cross-entropy loss.
  • Multi-Layer Perceptron (MLP): 3-layer neural network with manual backpropagation.
  • Predictor workflow for 28x28 PNG handwritten digits.

Stats Link to heading

  • Project page: /projects/mnist-from-scratch/
  • Primary language: Rust
  • Commits: 15
  • Created: 2026-05-05T01:11:12Z
  • Last updated: 2026-05-08T17:10:33Z