Visual forward-forward networks train convolutional neural networks without back-propagation

https://www.eurekalert.org/news-releases/1101843

"overcomes slow convergence, overfitting, high computational requirements, black box nature... training each network layer individually without back-propagation still faces information loss in input images, reducing classification accuracy, solved using: Label-Wise Noise Labeling preventing pixel information loss, Cosine Similarity-Based Contrastive Loss preserving spatial information for accurate image classification, Layer Grouping by output characteristics/ adding auxiliary layers, improving performance"