Adaptive Multi-Scale Goodness Aggregation Advances Forward-Forward Learning

A team of researchers has unveiled Adaptive Multi-Scale Goodness Aggregation (AMSGA), an innovative enhancement of the Forward-Forward (FF) algorithm tailored for local-learning neural networks. This new approach enhances stability, robustness, and generalization by integrating multi-scale goodness aggregation at local, intermediate, and global levels, employing adaptive curriculum-guided hard negative mining, layer-specific adaptive thresholds, and a warm-up cosine annealing learning-rate schedule. These enhancements maintain the memory-efficient and biologically plausible characteristics of the original FF model while overcoming its shortcomings. Testing on MNIST and Fashion-MNIST demonstrates significant performance gains over the baseline FF algorithm, with improvements of up to +1.45% on MNIST and +1.50% on Fashion-MNIST, all without notable computational costs. The findings are published in arXiv:2605.18804.