Learning
A real training run of the U-Net, as two waves that fall into step.
The full-length video is not published yet. This loop is a four-second excerpt.
What it shows
A neural network learning, as two waves that fall into step. The white wave is the answer to reach, and the coloured wave is what the network proposes at that point of its training: out of phase, noisy and red at first, then locked on and cyan. The gap between the waves is its error, and two balls show the error on training data and on validation data.
Data
The 67 training rounds (epochs) of the real run of the land-cover model published on Hugging Face (landscape-change-detection-unet): its errors and its accuracy score, the mIoU, which measures how well its prediction overlaps the truth.
How it is made
The target is three superimposed waves. The network’s wave uses the same ones, shifted and blurred in proportion to its error.
Sound
A reference note and the network’s note, more out of tune when the error is large, with some noise. A major chord appears as the accuracy rises, with a small tick per round.
© 2026 Maxime Tarka, all rights reserved. Made from the land-cover U-Net and its pipeline: model on Hugging Face, pipeline on GitHub.