Neural nets are now hunting ternary matrix-multiply schemes, and this one reports record-low addition counts.

Turns hard-label model extraction into inner-product algebra, ditching SVD clustering and extending to max-pooling CNNs.