Compact Hardware Synthesis of Stochastic Spiking Neural Networks.

Area of Science:

• Neuroscience and Artificial Intelligence
• Hardware Implementation of Neural Networks

Background:

• Spiking neural networks (SNNs) mimic biological neurons, offering high-fidelity emulation.
• Implementing high-density, biologically-inspired SNNs in hardware remains a significant scientific and technical challenge.

Purpose of the Study:

• To propose a compact digital design for high-volume SNNs that incorporates biological neuron stochasticity.
• To enhance processing speed and enable high-density hardware implementation of SNNs.

Main Methods:

• Developed a novel stochastic SNN (SSNN) model with a compact digital design.
• Compared the proposed SSNN model's performance against previous SSNN models and other SNNs using unsupervised STDP learning.
• Scaled the SSNN model for high-volume networks trained via backpropagation and applied it to pattern classification.

Main Results:

• The proposed SSNN model demonstrated superior processing speed compared to existing SSNN models.
• Achieved better results in pattern classification tasks compared to recently published SNN models using unsupervised STDP learning.
• The model's scalability for high-volume networks trained with backpropagation was successfully demonstrated.

Conclusions:

• The proposed compact digital SSNN design facilitates high-density, high-volume SNN hardware implementation.
• The SSNN model offers improved performance in terms of speed and accuracy for pattern classification tasks.
• This work contributes a scalable and efficient approach to realizing advanced SNNs in hardware.