Biologically plausible local synaptic learning rules robustly implement deep supervised learning.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Deep neural networks (DNNs) rely on middle-layer representational learning for efficiency.
• Current backpropagation (BP) learning rules lack biological plausibility for brain implementation.
• Developing brain-like learning rules is crucial for understanding memory and cognition.

Purpose of the Study:

• To develop and evaluate biologically plausible learning rules for DNNs.
• To compare these rules against established methods and animal learning performance.
• To investigate a novel variant, FA_Ex-100%, incorporating direct dopamine signaling.

Main Methods:

• Numerical simulations of DNNs performing a reward prediction task.
• Implementation and comparison of Extreme Learning Machine (ELM), Weight Perturbation (WP), and Feedback Alignment (FA) rules.
• Development and testing of FA_Ex-100% with simulated local error signals.

Main Results:

• FA achieved performance comparable to BP, outperforming ELM and WP.
• FA_Ex-100% performance was also comparable to BP.
• FA_Ex-100% demonstrated robustness against perturbations and noise.

Conclusions:

• Simplified, biologically plausible learning rules like FA_Ex-100% can support deep supervised learning.
• Accurate error signals, potentially via dopaminergic neurons, are key for robust learning.
• This research offers insights into brain-inspired AI and neural computation.