Deep Spike Learning With Local Classifiers.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep spiking neural networks (SNNs) offer computational efficiency but training them with backpropagation is resource-intensive and hinders parallelization.
• Biologically inspired local learning rules enable efficient training but typically achieve lower accuracy on practical tasks.
• A significant challenge remains in training deep SNNs effectively using local learning for both efficiency and high performance.

Purpose of the Study:

• To develop a supervised local learning scheme for training deep SNNs that achieves both computational efficiency and high accuracy.
• To propose novel spike-based local learning rules that consider temporal dependencies.
• To evaluate the proposed methods against established baselines on diverse datasets.

Main Methods:

• A supervised local learning scheme was employed, optimizing each layer independently with an auxiliary classifier.
• A novel spike-based efficient local learning rule was proposed, focusing on direct temporal dependencies.
• Two variants incorporating temporal dependencies via backward and forward processes were developed and tested.

Main Results:

• The proposed methods successfully scaled to large deep spiking neural network architectures.
• Experimental results demonstrated substantial performance improvements over spike-based local learning baselines across six datasets.
• Temporal dependencies were found crucial for high performance on temporal tasks, with minimal impact on rate-based tasks.

Conclusions:

• The developed local learning rules significantly advance the performance of spike-based local learning for deep SNNs.
• The methods retain the computational benefits of local learning while achieving competitive accuracy.
• This work bridges the gap between efficient and accurate training of deep SNNs using local learning strategies.