Deep learning in spiking neural networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep learning, particularly artificial neural networks (ANNs), has achieved remarkable success in machine learning and computer vision.
• ANNs utilize static neuron activations and require extensive labeled data for supervised training via backpropagation.
• Biological neurons communicate using discrete spikes, where timing and rate are crucial, making Spiking Neural Networks (SNNs) a more biologically plausible model.

Purpose of the Study:

• To review recent supervised and unsupervised methods for training deep Spiking Neural Networks (SNNs).
• To compare the accuracy and computational cost of different SNN training approaches.
• To assess the potential of SNNs as a more biologically realistic and efficient alternative to ANNs.

Main Methods:

• Review of recent supervised and unsupervised learning algorithms for deep SNNs.
• Comparative analysis of SNN training methods based on classification accuracy.
• Evaluation of computational efficiency, focusing on operational costs.

Main Results:

• Deep SNNs currently exhibit lower accuracy than ANNs on many tasks, but the performance gap is narrowing.
• SNNs demonstrate significantly lower computational costs compared to ANNs.
• SNNs show promise for processing spatio-temporal data due to their event-driven nature.

Conclusions:

• Training deep SNNs is challenging due to non-differentiable neuron activation functions, hindering backpropagation.
• SNNs offer a more biologically realistic and computationally efficient alternative to ANNs, especially for spatio-temporal data.
• While SNNs lag in accuracy, ongoing research is rapidly closing the gap, positioning them as a key area for future AI development.