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

Researchers introduce a learnable channel-class assignment mechanism for Forward-Forward (FF) neural networks, enabling adaptive specialization in convolutional layers. The method combines entropy and orthogonality regularization with loss-aware layer weighting to achieve state-of-the-art performance among FF-based models on image classification benchmarks, substantially narrowing the performance gap with traditional backpropagation.

This research addresses a fundamental challenge in biologically-inspired neural network training: the Forward-Forward algorithm provides an alternative to backpropagation that better aligns with biological neural computation, but existing implementations suffer from performance limitations on complex tasks. The core innovation—learnable channel-class assignment—moves beyond static architectural constraints that have constrained prior FF-based CNNs, allowing the network to dynamically specialize different channels for different classification tasks based on learned patterns rather than predetermined assignments.

The FF algorithm has gained theoretical interest because it eliminates the backward pass entirely, using only local forward-propagating objectives to guide learning. This property appeals to neuroscientists and researchers exploring more biologically plausible learning mechanisms. However, previous adaptations to convolutional architectures have struggled with performance gaps relative to standard backpropagation, limiting practical adoption. By introducing entropy and orthogonality regularization alongside adaptive layer contribution weighting, the authors create mechanisms that promote both diversity and stability in channel specialization.

The demonstrated performance improvements across CIFAR-10, CIFAR-100, and Tiny-ImageNet represent meaningful validation of the approach's effectiveness. Narrowing the performance gap with backpropagation is particularly significant because it strengthens the case for FF algorithms in scenarios where biological plausibility, hardware efficiency, or learning interpretability matter. For machine learning practitioners, this work suggests that forward-only learning remains viable for standard computer vision tasks when properly optimized.

Future research should explore whether these techniques extend to larger-scale datasets and modern architectures like Vision Transformers. The applicability of learnable channel assignment to other forward-only learning paradigms also warrants investigation.

• →Learnable channel-class assignment enables adaptive specialization in Forward-Forward networks, moving beyond static architectural constraints
• →Entropy and orthogonality regularization, combined with loss-aware layer weighting, significantly improve FF-based CNN performance
• →State-of-the-art FF results on CIFAR-10/100 and Tiny-ImageNet substantially reduce the performance gap with backpropagation
• →The approach validates that biologically-inspired forward-only learning is increasingly competitive for practical computer vision tasks
• →Future work should test scalability to larger datasets and modern architectures beyond residual CNNs