On-chip learning in about 2 ms

What the productive prototype path actually learns and what the timing means.

What the chip learns

48×48 two-channel binary silhouettes go into a FullyConnected layer with AkidaUnsupervised. Learning stores prototypes, not gradients.

Confidence

Confidence uses recovered pattern potential relative to num_weights, not a softmax share, so a clean part can exceed 90% without competing for probability mass with unrelated classes.

Three-card view distribution

Cards 0–2 learn the same objects but staggered subsets: card i gets every third view with an offset. The best match across the group wins.

Measured timing

~2 ms/example on AKD1500. One chip-learning dialog showed 79 examples / 3 objects completing under one second; step 3/3 took ~0.3 s.

Why recognized?

After learning, each source photo is replayed and attached to its strongest neuron. Runtime can show “Neuron 3 · Card 2” and the actual learned image beside the live crop.