ML — live coded autograd, four showcases

Pick a tab to swap models. Each showcase is a complete cljrs program — model, dataset, loss, training loop, all editable. Hold Alt (or Ctrl) and drag any number to scrub it live. The GPU toggle re-routes matmul to a wgpu compute shader on native; on the web it transparently falls back to CPU.

Stack: a tiny reverse-mode autograd over dense f32 tensors in cljrs-ml, exposed as builtins (ml/matmul, ml/matmul-gpu, ml/relu, ml/sigmoid, ml/softmax, ml/conv1d-valid, ml/cross-entropy, ml/adam-step!, ml/argmax, ml/one-hot, ...). Forward, backward, and optimizer step are all in Rust; everything else lives in the editor.

How the showcases work

• Function fit — a 1→H→1 MLP fits a noisy sin(2x). Loss is MSE, optimizer is plain SGD.
• 2D classifier — a two-layer MLP carves a decision boundary out of the two-moons dataset. Trained with cross-entropy + Adam; canvas paints the model's confidence underneath the points.
• MNIST-tiny — 8×8 hand-curated digits (200 train / 50 test, embedded in source) classified by a 64→32→10 MLP with cross-entropy + Adam. Stage shows test predictions; correct answers go green, mistakes go red.
• Autoencoder — a 64→16→64 net reconstructs the same 8×8 digits. Stage shows three side-by-side inputs and their reconstructions over time.

Use the GPU checkbox above to switch the showcase's matmul implementation between CPU and a wgpu compute kernel. Every showcase is built on a thin (matmul a b) macro that swaps based on a (USE-GPU?) atom JS sets each time you toggle.