Reward-optimizing learning using stochastic release plasticity

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Synaptic plasticity enables adaptive learning in neural systems and is a biologically plausible model for reward-driven learning.
• A key challenge is developing plasticity rules that match the robustness and effectiveness of error backpropagation.

Purpose of the Study:

• Introduce Reward-Optimized Stochastic Release Plasticity (RSRP), a new learning framework.
• Derive a plasticity rule that maximizes reward signals using natural gradient estimation.
• Evaluate RSRP's performance and stability in reinforcement learning and digit classification tasks.

Main Methods:

• Model synaptic release as a parameterized distribution within the RSRP framework.
• Employ natural gradient estimation to derive the RSRP learning rule.
• Validate RSRP in biologically plausible neural networks and compare it with Proximal Policy Optimization (PPO) and error backpropagation.

Main Results:

• RSRP demonstrates competitive performance and stability in reinforcement learning, on par with PPO.
• RSRP achieves accuracy comparable to error backpropagation in digit classification tasks.
• Reward regularization is identified as a crucial mechanism for stabilizing RSRP.

Conclusions:

• RSRP provides a robust and effective synaptic plasticity learning rule.
• The findings have implications for both artificial intelligence and experimental neuroscience, particularly in discontinuous reinforcement learning scenarios.