Spiking CMOS-NVM mixed-signal neuromorphic ConvNet with circuit- and training-optimized temporal subsampling

Area of Science:

• Neuromorphic Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are crucial for visual data processing but are energy-intensive when deployed on conventional hardware.
• Spiking neural networks (SNNs) and neuromorphic hardware offer energy-efficient alternatives, but converting DNN architectures like Convolutional Neural Networks (CNNs) is challenging.
• A key hurdle is the lack of efficient spiking implementations for CNN functions, such as MaxPool, hindering neuromorphic hardware development.

Purpose of the Study:

• To propose and evaluate a novel MaxPool layer with temporal multiplexing for Spiking CNNs (SCNNs).
• To enable the implementation of SCNNs on energy-efficient, mixed-signal neuromorphic hardware.
• To address the challenge of converting DNN functionality to spiking domain with minimal accuracy loss.

Main Methods:

• Leveraging the temporal dynamics of Integrate & Fire neurons for MaxPool decision-making in the spiking domain.
• Implementing and testing the proposed MaxPool in an SCNN architecture using the aihwkit framework.
• Proposing a hardware realization of the spiking MaxPool in a non-volatile memory (NVM) crossbar array with CMOS periphery circuits.

Main Results:

• The proposed spiking MaxPool achieves up to 61.74% accuracy on CIFAR10 with 10% spatiotemporal input, showing a minimal 1% drop compared to 62.78% with 100% input (2-bit weight resolution).
• The scheme allows selective trade-off between latency and accuracy, enabling decisions before full spatiotemporal input is processed.
• Energy-efficiency estimations indicate competitive performance against existing neuromorphic chip designs.

Conclusions:

• The proposed temporal multiplexing MaxPool is amenable for mixed-signal circuit implementation on neuromorphic hardware.
• This work facilitates the conversion of CNNs to energy-efficient SCNNs, advancing neuromorphic computing.
• The developed technique offers a promising approach for low-power, high-accuracy visual data processing on specialized hardware.