ReStoCNet: Residual Stochastic Binary Convolutional Spiking Neural Network for Memory-Efficient Neuromorphic...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer energy-efficient computation.
• Deep SNNs face challenges in hierarchical feature learning and memory footprint.
• Binary kernels can reduce SNN memory but may impact performance.

Purpose of the Study:

• To propose ReStoCNet, a novel residual stochastic multilayer convolutional SNN with binary kernels.
• To enhance computational efficiency and reduce synaptic memory in SNNs.
• To improve deep SNNs' hierarchical feature learning capabilities for complex pattern recognition.

Main Methods:

• Developed ReStoCNet with stacked convolutional layers, pooling, and fully-connected layers.
• Incorporated residual connections to enhance deep SNN feature learning.
• Utilized a novel Hybrid-STDP (HB-STDP) unsupervised learning algorithm for binary kernels.

Main Results:

• ReStoCNet demonstrated effective hierarchical feature extraction and inference.
• Residual connections mitigated accuracy loss in deep SNNs.
• Achieved over 20x kernel memory compression compared to full-precision SNNs on MNIST and CIFAR-10 datasets.

Conclusions:

• ReStoCNet and HB-STDP provide an efficient and accurate SNN solution for pattern recognition.
• Residual connections are crucial for deep SNN performance.
• Binary kernels significantly reduce memory footprint without compromising classification accuracy.