Exponentiated gradient learning yields brain-like synaptic distributions.

Exponentiated gradient learning produces synaptic weight distributions that closely resemble brain-like log-normal patterns and maintains synaptic sign stability.

• Why it matters: Understanding how synaptic distributions form and function is crucial for modeling brain learning accurately, as traditional gradient descent methods fail to replicate these biological features.
• What they did: The study applied exponentiated gradient (EG) learning to recurrent neural networks, demonstrating that it naturally enforces stable synaptic signs and generates realistic weight distributions, while preserving task performance.
• The result: EG-trained networks show increased robustness to synaptic pruning and improved learning from sparse inputs, linking an optimization principle to observed neural organization and advancing biologically plausible models.