Sharing leaky-integrate-and-fire neurons for memory-efficient spiking neural networks.

Area of Science:

• Artificial Intelligence
• Computer Science
• Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer energy-efficient computation through binary and asynchronous operations.
• Leaky-Integrate-and-Fire (LIF) neurons, crucial for SNNs, require substantial memory to store membrane voltage for temporal dynamics.
• Memory requirements for LIF neurons escalate with larger input dimensions, posing a challenge for SNN scalability.

Purpose of the Study:

• To introduce a novel technique for reducing memory consumption in SNNs, specifically addressing the memory demands of LIF neurons.
• To develop an efficient SNN architecture that maintains high accuracy while optimizing memory usage.

Main Methods:

• Proposed EfficientLIF-Net, a novel SNN architecture that enables sharing of LIF neurons across different layers and channels.
• Implemented and evaluated the EfficientLIF-Net on diverse benchmark datasets including CIFAR10, CIFAR100, TinyImageNet, ImageNet-100, and N-Caltech101.
• Assessed the performance of EfficientLIF-Net on Human Activity Recognition (HAR) datasets, emphasizing its utility in temporal information processing.

Main Results:

• Achieved comparable accuracy to standard SNNs.
• Demonstrated significant memory efficiency gains: up to ~4.3x forward and ~21.9x backward memory efficiency for LIF neurons.
• Validated the effectiveness of EfficientLIF-Net across various image classification and HAR tasks.

Conclusions:

• EfficientLIF-Net presents a simple yet effective solution for enhancing memory efficiency in SNNs without compromising accuracy.
• The proposed neuron-sharing strategy offers substantial memory savings, making SNNs more practical for large-scale applications.
• EfficientLIF-Net shows promise for applications requiring efficient temporal information processing, such as HAR.