Achieving Optimal Accuracy and Robustness Through Tight Excitatory-Inhibitory Balance in Shallow Spiking Recurrent...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional deep neural networks (DNNs) face challenges with high computational complexity and lack of biological interpretability.
• Spiking Recurrent Neural Networks (SRNNs) offer biological plausibility and efficiency in processing spatio-temporal data using discrete spike events.

Purpose of the Study:

• To propose and evaluate an excitation-inhibition balanced shallow SRNN (EI-SRNN) for improved performance.
• To investigate the impact of balanced excitation and inhibition on SRNN accuracy, robustness, and computational complexity.

Main Methods:

• Developed an EI-SRNN by optimizing reservoir neuron input currents to achieve a tight balanced state, inspired by brain neurodynamics.
• Analyzed neural encoding ability and information memory capacity of the EI-SRNN.
• Compared model performance under varying degrees of excitation and inhibition.

Main Results:

• The EI-SRNN achieved optimal accuracy with low computational complexity, challenging the accuracy-robustness trade-off.
• Tight balanced excitatory and inhibitory states resulted in higher neural coding and memory capacity.
• Performance degraded more rapidly when the reservoir was dominated by excitation compared to inhibition.

Conclusions:

• The EI-SRNN demonstrates superior accuracy and robustness by leveraging balanced excitation and inhibition.
• Optimizing for balanced states enhances neural coding and memory capacities in SRNNs.
• EI-SRNN presents a biologically plausible and computationally efficient alternative to traditional DNNs.