Backpropagation-free spiking neural networks with the forward-forward algorithm.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) mimic biological neurons using discrete spikes, offering potential efficiency.
• Traditional backpropagation (BP) training for SNNs faces challenges in computational cost and biological realism.

Purpose of the Study:

• To investigate the Forward-Forward (FF) algorithm as an alternative training method for SNNs.
• To evaluate the performance of an FF-based SNN framework on various datasets.

Main Methods:

• Implemented a novel FF-based training framework for SNNs.
• Evaluated the framework on static (MNIST, Fashion-MNIST, Kuzushiji-MNIST) and spiking (Neuro-MNIST, SHD) datasets.
• Compared performance against existing FF-based SNNs and backpropagation-trained SNNs.

Main Results:

• The FF-based SNN model achieved superior performance over prior FF-SNNs on static datasets with a lighter architecture.
• Accuracy comparable to state-of-the-art BP-trained SNNs was reached on static datasets.
• Outperformed other SNN models on complex spiking tasks (SHD) and remained competitive with BP-SNNs.

Conclusions:

• The FF algorithm presents a promising alternative for SNN training, enhancing efficiency and biological plausibility.
• This approach addresses key limitations associated with backpropagation in SNNs.
• FF-based SNNs show potential for advancing neuromorphic computing and AI applications.