MNIST Playground

An interactive in-browser MNIST lab with ONNX Runtime inference, drawing, preprocessing traces, feature heatmaps, and linked embedding/logit spaces.

Status: active · GitHub · Demo


What it is

A static SvelteKit lab for exploring MNIST classifiers directly in the browser. It packages trained PyTorch models as ONNX artifacts, then lets users draw digits, inspect normalized tensors, compare probabilities/logits, and see how convolutional layers respond through feature heatmaps.

Why it matters

The app turns a small CNN into a visible system: model variants, training perturbations, UMAP-style embeddings, logit space, confusion analysis, and validation proof are all shipped as static artifacts. The free-draw robust model is trained against hard cases like mirrored 3s and 7s, rotated digits, and upside-down-ish 1s while preserving normal MNIST accuracy.

Demo · GitHub