A compressed sensing neuromorphic processor for sparse signal classification.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation for AI tasks.
• Sparse signal classification requires efficient processing to manage high dimensionality.
• Existing compression methods for SNNs often involve trade-offs between compression ratio and accuracy.

Purpose of the Study:

• To develop an integrated neuromorphic processing system for sparse signal classification.
• To enhance SNNs with compressed sensing for improved efficiency and model compression.
• To design and validate a specialized hardware processor for the proposed CSSNN.

Main Methods:

• Integration of data coding, compression, and SNN classification into a unified Compressed Sensing Spiking Neural Network (CSSNN).
• Evaluation of CSSNN on MNIST, N-MNIST, and DVS Gesture datasets with varying compression ratios (CRs).
• Design of a spike-driven CSSNN processor and validation on FPGAs and 40 nm CMOS process for ASIC.

Main Results:

• CSSNN achieved over 80% reduction in network operations compared to fixed Gaussian random matrix methods with minimal accuracy loss.
• The ASIC implementation demonstrated 96.12% classification accuracy on MNIST with 8-bit quantization.
• Estimated energy consumption of the ASIC is 2.089 mW at 100 MHz and 1.1-V supply.

Conclusions:

• The proposed CSSNN provides an effective approach for efficient sparse signal classification in neuromorphic systems.
• The specialized CSSNN processor enables high performance and low power consumption for real-world applications.
• End-to-end optimization of network performance and model compression is achievable with the integrated CSSNN framework.