🎧 Listen to a Neural Net Learn

A tiny neural network learns to untangle two spirals — and its training is the score. It starts as an out-of-tune orchestra and tunes itself as it learns. Headphones on, then press play.

fresh model — silence before the music

problem

learning rate

volume

The drone chord has one voice per layer, detuned by how hard that layer is being corrected. The melody plays one note per classified point — in-scale when the net is right, a sour semitone off when it's wrong, louder for bigger mistakes. When the chord settles pure and the melody goes quiet and consonant, it has learned.