Correlative Information Maximization: A Biologically Plausible Approach to Supervised Deep Neural Networks without Weight Symmetry

A team of researchers including Bariscan Bozkurt, Cengiz Pehlevan, and Alper T. Erdogan has published a significant paper titled 'Correlative Information Maximization: A Biologically Plausible Approach to Supervised Deep Neural Networks without Weight Symmetry' on arXiv. The paper addresses a fundamental criticism of current AI systems—their biological implausibility—by proposing an alternative to the backpropagation algorithm that has dominated deep learning. The researchers' Correlative Information Maximization (CIM) framework offers a normative approach to signal propagation in both forward and backward directions within neural networks, creating structures that more closely resemble biological systems with multi-compartment pyramidal neurons and lateral inhibitory neurons.

The key innovation lies in how CIM solves the long-standing 'weight symmetry problem' that has plagued backpropagation's biological plausibility. In conventional backpropagation, forward and backward paths require symmetric weights—a feature not observed in biological brains. CIM leverages two alternative yet mathematically equivalent forms of the correlative mutual information objective, which intrinsically generates separate forward and backward prediction networks without requiring weight symmetry. This breakthrough, combined with coordinate descent optimization and mean square error loss functions, provides a more realistic model of how biological neural networks might perform supervised learning tasks while maintaining computational efficiency.

The research, originally submitted in June 2023 and recently revised in March 2026, represents a significant step toward bridging the gap between artificial and biological intelligence. By moving away from backpropagation's biologically implausible requirements while maintaining supervised learning capabilities, the CIM framework opens new avenues for neuromorphic computing and brain-inspired AI architectures that could lead to more energy-efficient and robust learning systems.