Beyond Backpropagation — Brain-Inspired Learning for Vision

Modern vision performance is powered by backpropagation, yet biological learning suggests a different template: local credit assignment modulated by global signals. This project develops a backprop-free training regime in which parameter updates emerge from decentralized agents embedded in a network, coordinated through neuromodulatory rewards. Framed as continuous control in synaptic space, updates are treated as real-valued actions; local learners adjust weights using low-variance credit assignment inspired by MAP Propagation and coagent frameworks.

Compact ConvNets or patch-token transformers serve as testbeds for studying stability, sample-efficiency, and robustness under noisy or sparse feedback. The investigation probes which inductive biases—layerwise credit routing, temperature schedules, or exploratory perturbations—most closely approximate gradient transport while retaining biological plausibility. Emphasis is placed on calibration, corruption robustness, and compute efficiency, aiming to show that competitive learning dynamics arise without explicit backpropagation, with behavior that feels closer to cortex than calculus.

References:
https://stephen-c.com/projects/coagent/

Chung, Stephen. "Map propagation algorithm: Faster learning with a team of reinforcement learning agents." Advances in Neural Information Processing Systems 34 (2021): 14734-14744.