Adaptive Multi-Scale Goodness Aggregation for Forward-Forward Learning
The paper introduces Adaptive Multi-Scale Goodness Aggregation (AMSGA), an enhancement of the Forward-Forward algorithm aimed at improving neural network performance. AMSGAs modifications include multi-scale goodness aggregation and adaptive learning techniques, leading to better stability and generalization. Experimental results show significant performance improvements on standard datasets without added computational costs.
- ▪AMSGA is designed to enhance the stability and robustness of local-learning neural networks.
- ▪The proposed method includes adaptive curriculum-guided hard negative mining and layer-dependent adaptive thresholds.
- ▪Experiments on MNIST and Fashion-MNIST show performance improvements of up to +1.45% and +1.50%, respectively.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.18804 (cs) [Submitted on 11 May 2026] Title:Adaptive Multi-Scale Goodness Aggregation for Forward-Forward Learning Authors:Salar Beigzad, Vansh Verma View a PDF of the paper titled Adaptive Multi-Scale Goodness Aggregation for Forward-Forward Learning, by Salar Beigzad and Vansh Verma View PDF HTML (experimental) Abstract:We propose Adaptive Multi-Scale Goodness Aggregation (AMSGA), a novel extension of the Forward-Forward (FF) algorithm designed to improve stability, robustness, and generalization in local-learning neural networks.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
