Reproduce “Local plasticity rules can learn deep representations using self-supervised contrastive predictions” on OpenResearch

About this paper

Bernd Illing, Jean Ventura, Guillaume Bellec, Wulfram Gerstner

The research introduces CLAPP, a biologically plausible learning rule, demonstrating how deep neural networks can learn hierarchical representations through self-supervised contrastive predictions based on temporal cues like saccades. This approach, which integrates dendritic computation and local plasticity, achieved 73.6% accuracy on STL-10 image classification by leveraging its learned features.

Key insights

• Intrinsic temporal dynamics, such as self-generated gaze shifts, can provide sufficient self-supervision for deep representation learning under biological constraints.
• Predictive dendritic input can act as a "third factor" to modulate local synaptic plasticity, enabling effective credit assignment in deep neural networks.
• Deep, hierarchical representations are achievable using strictly local, online plasticity rules, demonstrating that biological plausibility does not inherently prevent complex feature learning.

Summary by alphaXiv.