What is a neural network?
A picture becomes numbers. Small adjustments become recognition.
Watch a network learn to see.
A handwritten eight passes through a luminous network. Connections carry different amounts of influence; errors guide small adjustments. The sculpture stays in motion as the explanation follows the mathematics.
Input brightness follows pixel values, hidden light follows computed activity, and output light follows model scores. Light travels backward during the gradient passage. Camera motion, curved fibers, and atmosphere give the calculation a spatial form.
Look closer: the mathematics
A 64 → 12 ReLU → 10 softmax network, with 888 weights and 22 biases, trained on 1,438 handwritten images. Across 359 held-out images it improves from 41 to 351 correct, leaving eight mistakes. The split is by image, not by independent writer. The film interpolates recorded training parameters; final predictions hold all parameters fixed. Softmax scores are not calibrated probabilities of correctness. Fiber shape, travel speed, and lighting are artistic choices.
Sources
How the work is described ↗- Made
- Oct 5, 2026
- Study
- Full film · Thyone score
- Sound
- Thyone — Ben Elson
Ben Elson / Thyone / courtesy of Epidemic Sound