Noise and Dynamical Synapses as Optimization Tools for Spiking Neural Networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Standard artificial neural networks (ANNs) exhibit limited flexibility with corrupted inputs due to their fixed architecture.
• Biological neural systems demonstrate robustness and adaptability through temporal coding mechanisms.

Purpose of the Study:

• To investigate the use of biological temporal coding features in spiking neural networks (SNNs) to enhance performance with non-optimal parameters.
• To demonstrate the advantages of SNNs over ANNs in handling complex, linearly inseparable problems.

Main Methods:

• Utilized noise-induced stochastic resonance and dynamical synapses within a spiking neural network model.
• Employed the analog XOR task as a simplified convolutional neural network (CNN) model to evaluate performance.
• Compared the efficacy of SNNs against traditional ANNs.

Main Results:

• Spiking neural networks (SNNs) successfully solved the linearly inseparable analog XOR problem using fewer neurons compared to ANNs.
• In leaky SNNs, the integration of noise and dynamical synapses compensated for non-optimal parameters, yielding near-optimal results for weaker inputs.

Conclusions:

• SNNs present a more flexible and efficient alternative to ANNs, particularly for tasks with noisy or imperfect input data.
• Biological temporal coding mechanisms like stochastic resonance and dynamical synapses are effective strategies for improving SNN robustness and performance without parameter optimization.