Train a digit classifier on your GPU

A 64→64→10 MLP learning UCI's 8×8 handwritten digits with Adam. Every step — forward pass, loss, backward pass and the optimizer's update — is one graph in this project's IR, compiled to WGSL kernels by its own compiler and run in your browser. The grey curve is PyTorch's run from the same initial weights on the same batches, in float64; yours is float32 on your GPU. The report says what was checked and how.

Data: UCI Optical Recognition of Handwritten Digits (Alpaydin & Kaynak, 1998; doi:10.24432/C50P49), CC BY 4.0, served from this repository and checked against its SHA-256 before use. Nothing here is timed. The power mode of the machine is not readable from a browser; it is recorded only when this page runs on the project's own local server.