Dev builds NumPy-only MLP trainer with live visualizations of weights, t-SNE and ablation

No-Brain-1655 · reddit · 2026-09-27

A developer open-sourced an educational tool implementing a small MLP in pure NumPy (manual backprop, SGD+momentum, L2, dropout, cosine decay) reaching 98.5% on MNIST. The GUI shows live loss curves, per-layer gradient norms, inactive neuron ratios, weight distributions vs init, first-layer receptive fields, per-layer PCA/t-SNE with misprediction links, noise/rotation robustness curves, and an interactive lab for ablating neurons, pruning, adding weight noise, and tuning softmax temperature with instant accuracy updates. Aimed at students and teachers from high school to intro ML courses.

Original post →

More from coding & agent

coding & agent channel →