An adaptive threshold neuron for recurrent spiking neural networks with nanodevice hardware implementation

Area of Science:

• Neuromorphic Engineering
• Artificial Neural Networks
• Device Physics

Background:

• Recurrent Spiking Neural Networks (RSNNs) are crucial for processing temporal data but face challenges in convergence speed and accuracy.
• Existing neuron models often lack the flexibility required for complex tasks and efficient hardware implementation.
• Advancements in non-filamentary resistive switching devices offer potential for energy-efficient neuromorphic hardware.

Purpose of the Study:

• To propose and validate a new Double EXponential Adaptive Threshold (DEXAT) neuron model for improved RSNN performance.
• To develop a hardware-efficient realization methodology for DEXAT neurons using circuit-device co-design.
• To demonstrate the practical application of DEXAT-based RSNNs in real-time speech recognition tasks.

Main Methods:

• Development of the DEXAT neuron model incorporating adaptive threshold dynamics.
• Design and fabrication of a hardware-efficient neuron block using oxide-based non-filamentary resistive switching devices.
• Simulation of a full RSNN using experimentally extracted device parameters for performance evaluation.
• Experimental demonstration of end-to-end real-time speech recognition inference using fabricated hardware.

Main Results:

• The DEXAT neuron model achieved faster convergence and higher accuracy compared to conventional models.
• Simulated RSNNs attained 96.1% accuracy on the SMNIST dataset and 91% on the Google Speech Commands (GSC) dataset.
• Real-time speech recognition was successfully demonstrated using fabricated resistive memory circuit-based DEXAT neurons.
• Investigations confirmed the robustness of DEXAT-based RSNNs against nanodevice variability and endurance issues.

Conclusions:

• The proposed DEXAT neuron model significantly enhances the performance of neuromorphic RSNNs.
• The hardware-efficient implementation using resistive switching devices enables practical, low-power neuromorphic systems.
• DEXAT-based RSNNs show strong potential for real-world applications like speech recognition, exhibiting robustness and efficiency.