Elsewhere

A pattern becomes a prediction.

Each luminous sheet is a feature map computed by a small convolutional neural network. Brightness and relief show the strength of its response.

1. Input A generated ring, wave or cross.
2. Convolution Learned 3 × 3 filters scan local patches.
3. Depth Three layers combine responses, using ReLU and stride 2.
4. Prediction Global average pooling and softmax produce three scores.

An independently trained teaching network using synthetic shapes. Inference runs locally; learning is not happening live. Maps are normalised for display, and the travelling light guides the eye rather than measuring execution time.