Skip to main content

Weight Sharing

This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.

State

A convolutional layer uses the same template weights at every position — a pattern learned anywhere is recognized everywhere, and the knob count collapses to template-sized.

Show

The corner cat and the centered cat finally meet the same detectors; the overfitting appetite drops to a feedable size.

Watch for

Weight sharing made deep seeing learnable at all — the one big trick, not an optimization detail.