Homeostatic Visual Cortex

A neural network that learns to reconstruct your webcam feed in real time. Every frame is a single forward+backward pass — there are no epochs, no batches, no pre-training. The model starts from random weights and learns continuously.

Reality Your live webcam feed, resized to the current working resolution.

Dream The autoencoder's reconstruction. Watch it sharpen from noise as the network learns.

Affect A limbic system inspired by MicroPsi theory. Two hydraulic tanks modulate learning: Competence fills when loss improves, drains during stagnation. Certainty fills when error is low, drains when surprised. High arousal (1 − certainty) increases the learning rate. If competence crashes while arousal spikes, TUNNEL VISION engages — dropping resolution to 64×64 so the network can recover faster.

U-Net autoencoder (3→16→32→64→128) with skip connections, ~380K parameters. Running via TensorFlow.js Core API on your GPU. No data leaves your browser.