Forward-Only Convolutional Neural Networks with Learnable Channel-Class Assignment

The paper presents a novel approach to Forward-Forward (FF) convolutional neural networks by introducing a learnable channel-class assignment mechanism that allows for adaptive specialization of convolutional channels. This method incorporates entropy and orthogonality regularization, along with a loss-aware layer contribution strategy, resulting in improved performance across benchmark datasets such as CIFAR-10, CIFAR-100, and Tiny-ImageNet. The proposed architecture not only achieves state-of-the-art results among FF-based models but also significantly reduces the performance gap with traditional backpropagation methods, highlighting its potential for enhancing representational capacity in deep learning applications.