Multi-compartment neuron and population encoding powered spiking neural network for deep distributional reinforcement...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) mimic brain processing for energy efficiency and biological realism.
• Existing SNN models often simplify neurons, limiting computational and learning capabilities.
• The leaky integrate-and-fire (LIF) model is common but neglects neuron structure.

Purpose of the Study:

• To develop a brain-inspired deep reinforcement learning algorithm using SNNs.
• To integrate a biologically realistic multi-compartment neuron (MCN) model into SNNs.
• To enhance computational power and learning in SNNs by incorporating structural neuron properties.

Main Methods:

• Proposed a deep distributional reinforcement learning algorithm based on SNNs.
• Integrated a bio-inspired multi-compartment neuron (MCN) model simulating dendritic and somatic compartments.
• Introduced an implicit fractional embedding method using population coding of spiking neurons.

Main Results:

• The proposed model, MCS-FQF, outperformed vanilla FQF (ANN-based) and Spiking-FQF (ANN-to-SNN conversion) on Atari games.
• Ablation studies confirmed the benefits of the MCN model and the population spike representation.
• The novel SNN approach demonstrated enhanced performance and reduced power consumption.

Conclusions:

• The multi-compartment neuron model significantly boosts SNN computational power and learning.
• Population coding with implicit fractional representation enhances SNN performance and efficiency.
• This brain-inspired SNN approach offers a promising direction for advanced AI and neuromorphic computing.