Biological Plausibility and Representational Alignment of Feedback Alignment in Convolutional Networks

Researchers demonstrate that modified feedback alignment (FA) algorithms can train convolutional neural networks while maintaining biological plausibility, with internal representations converging to structures similar to backpropagation despite using fundamentally different weight update mechanisms. This finding suggests that successful learning algorithms may achieve comparable results through different computational paths, bridging biologically plausible alternatives with practical neural network training.

This research addresses a fundamental challenge in neuroscience-inspired machine learning: developing algorithms that are both biologically realistic and practically effective. Feedback alignment emerged as a theoretically appealing alternative to backpropagation because biological neural systems cannot implement backpropagation's weight transport problem, yet standard FA fails to scale to convolutional architectures commonly used in modern deep learning. The comparative analysis across five learning algorithms reveals an unexpected convergence—modified FA variants develop internal representational geometries nearly identical to backpropagation despite relying on random feedback connections rather than precise error gradients. This finding has significant implications for understanding learning principles. It suggests that achieving particular representational structures may be a fundamental attractor for diverse learning rules, implying that multiple computational mechanisms can solve the same learning problem. For the broader machine learning community, this work strengthens the case for biologically inspired learning algorithms by demonstrating functional equivalence with conventional approaches. The research demonstrates that computational simplicity and biological plausibility need not be sacrificed for performance on standard benchmarks. However, the practical impact remains limited—backpropagation remains more efficient and scalable. The neuromorphic computing community benefits most directly, as these findings validate pursuing biologically plausible training rules for specialized hardware and brain-inspired architectures where backpropagation implementation is challenging.

• →Modified feedback alignment converges on representational structures similar to backpropagation despite using different weight update mechanisms.
• →Biological plausibility can be maintained in neural network training without sacrificing performance on standard benchmarks.
• →Multiple fundamentally different learning algorithms appear to achieve similar internal representations, suggesting convergence to optimal representational geometry.
• →The research supports development of neuromorphic computing systems using biologically inspired learning rules.
• →Success in machine learning may depend more on achieving correct representational structures than on the specific computational path taken.