Hyperspherical Forward-Forward with Prototypical Representations
We fix Forward-Forward's slow inference by replacing per-class passes with a single forward pass through hyperspherical prototype matching. Thus, we achieve 40× faster inference with competitive accuracy.
