Researchers boost Forward-Forward learning efficiency by 22.6% using selective neuron measurement instead of traditional sum-of-squares methods

Forward-Forward algorithm offers a biologically plausible alternative to backpropagation by training neural networks layer by layer with local goodness functions

Top-k goodness function measures only the most active neurons and substantially outperforms the traditional sum-of-squares (SoS) approach on Fashion-MNIST

Entmax-weighted energy method adds learnable sparse weighting based on alpha-entmax transformation for additional performance improvements

Separate label feature forwarding (FFCL) technique injects class hypotheses at every layer to further enhance learning