Training Convolutional Neural Networks with the Forward-Forward algorithm

• DOI: 10.48550/arxiv.2312.14924
• Publication Date: 2023-01-01
• Authors: Riccardo Scodellaro; Ajinkya Kulkarni; Frauke Alves; Matthias Schröter
• Keywords: MNIST database; Backpropagation; Hyperparameter; Rprop

The recent successes in analyzing images with deep neural networks are almost exclusively achieved Convolutional Neural Networks (CNNs). training of these CNNs, and fact all network architectures, uses the backpropagation algorithm where output is compared desired result difference then used to tune weights towards outcome. In a 2022 preprint, Geoffrey Hinton suggested an alternative way which passes results together at input network. This so called Forward (FF) has up now only been fully connected networks. this paper, we show how FF paradigm can be extended CNNs. Our FF-trained CNN, featuring novel spatially-extended labeling technique, achieves classification accuracy 99.16% on MNIST hand-written digits dataset. We different hyperparameters affect performance proposed compare CNN trained standard approach. Furthermore, use Class Activation Maps investigate type features learnt by algorithm.