Spike-based alignment learning solves the weight transport problem.

Area of Science:

• Computational neuroscience
• Machine learning
• Artificial intelligence

Background:

• Traditional learning algorithms face challenges with symmetry constraints in physical neuronal networks.
• Contrastive Hebbian learning requires symmetric connectivity, and error backpropagation needs forward weights for backward passes, creating a weight transport problem.

Purpose of the Study:

• To introduce spike-based alignment learning (SAL), a synapse-local learning rule designed to solve the weight transport problem.
• To enable local computation and maintain gradient representation in neuronal networks.

Main Methods:

• Developed SAL, a learning rule that utilizes noise for synapse-local weight alignment.
• Demonstrated SAL's ability to operate alongside other learning rules to preserve symmetry.
• Validated SAL's efficacy in alleviating discrepancies from neuron and synapse variability in analog substrates.

Main Results:

• SAL successfully addresses the weight transport problem in neuronal network learning.
• The rule maintains necessary symmetry throughout learning, ensuring accurate local gradient representation.
• SAL demonstrated effectiveness in spiking Bayesian inference, bio-plausible error backpropagation, and a deep learning computer vision task.

Conclusions:

• SAL offers a novel, synapse-local solution for weight transport in neuronal networks.
• This learning rule is robust to variability in analog substrates, paving the way for more efficient neuromorphic computing.
• SAL enhances the feasibility of implementing complex learning algorithms in physical neuronal systems.