Neuroevolution Guided Hybrid Spiking Neural Network Training.

Area of Science:

• Neuromorphic computing
• Artificial intelligence
• Computational neuroscience

Background:

• Spiking Neural Networks (SNNs) are a key area in machine learning research.
• Current SNN training methods often lack optimization for SNN-specific properties.
• There is a need for adaptive and explainable neuromorphic algorithms.

Purpose of the Study:

• To develop a structured algorithmic framework for SNN training.
• To optimize unique SNN properties, such as neuron spiking thresholds, using neuroevolution.
• To create a feedback-based learning strategy for adaptable neuromorphic systems.

Main Methods:

• Utilizing neuroevolution as a feedback strategy for SNN training.
• Developing a hybrid bio-inspired training approach.
• Evaluating the algorithm on image classification datasets (CIFAR-10, CIFAR-100, ImageNet).

Main Results:

• Achieved significant latency improvements: 53.8% on CIFAR-10, 28.8% on CIFAR-100, and 28.2% on ImageNet compared to conversion-based methods.
• Demonstrated latency improvements of 43.2% on CIFAR-10 and 27.9% on CIFAR-100 for adversarial attack scenarios.
• Showcased the adaptability and explainability of the neuroevolution-based SNN training.

Conclusions:

• The proposed neuroevolutionary framework offers a powerful method for training SNNs.
• This approach leads to more efficient and adaptable neuromorphic systems.
• The algorithm shows promise for applications including adversarial image classification.