Covariance-Aware Goodness for Scalable Forward-Forward Learning

Researchers propose Bi-axis Covariance Goodness (BiCovG), a drop-in goodness function replacement for Hinton's Forward-Forward algorithm that captures second-order feature dependencies via cross-channel covariance projections and multi-scale spatial correlation statistics. Standard sum-of-squares goodness collapses feature volumes into scalar activation energies, losing inter-feature structure critical on ImageNet-100 and Tiny-ImageNet. BiCovG approximates full covariance without O(C²) overhead. Engineers exploring backpropagation-free training for memory-constrained or neuromorphic hardware deployments should evaluate this immediately.