[andre15silva]

Adds my dashboard cell and turns it on in cells/manifest.json.

What it shows

Four tiny neural networks race to learn a painting by Maria Helena Vieira da Silva from scratch, live in the browser. The paintings alternate: Bibliothèque (1949), then L'arène (1950), one minute per race. Each network maps a pixel's position (x, y) to its colour:

• ReLU MLP: the plain baseline
• SIREN: sine activations
• Fourier features: random sinusoid inputs
• Hash grid: multiresolution, Instant-NGP style

The race is fair, and the caption says so:

This is a race between four tiny neural networks with the same theoretical capacity learning the same painting with the same compute.

• Same capacity: each network has exactly 8,707 trainable float32 parameters. The cell checks this at startup and prints the result in the log.
• Same compute: each network trains in its own Web Worker and gets exactly the same CPU time (0.5 CPU-s per second).
• Tuning: the hyperparameters were tuned briefly per network, so none is handicapped.

Layout

• Fullscreen: the target next to each network's live guess with its PSNR, a leaderboard (overtakes are logged), PSNR curves, and a tail -f train.log with model summaries and step, loss, PSNR, learning rate, gradient norm and it/s.
• End of each race: the winner's snapshots next to the target, while the lanes alternate between each network's final guess and its error map.
• Grid tile: only the leading network and one live line.

Notes

• Self-contained: the training engine is hand-written JS with no dependencies. It falls back to running on the page if Web Workers are unavailable.
• Pauses when hidden: training stops while the tab is hidden.
• Restarts on fullscreen: a fresh race starts when the cell goes fullscreen in monitor mode, so viewers see it from noise.
• Embedded paintings: paintings.js embeds 160×128 RGB copies of the two paintings (164 KB). The network needs their pixels, and the museum images can't be read cross-origin.
• Copyright: Vieira da Silva's works are still under copyright.

Testing

Checked in headless Chromium:

• standalone at 1920×1080
• inside the board's sandboxed iframe, where the Web Workers run fine
• through the switch from one race to the next

🤖 Generated with Claude Code