Proposal for an All-Spin Artificial Neural Network: Emulating Neural and Synaptic Functionalities Through Domain Wall...

Area of Science:

• Neuromorphic Engineering
• Spintronics
• Artificial Intelligence

Background:

• Emerging post-CMOS technologies offer potential for low-power neural computing.
• Existing neuromorphic architectures often focus on mimicking either neurons or synapses individually.
• Spintronic devices show promise for efficient neuron thresholding, while memristive devices are explored for synapse emulation.

Purpose of the Study:

• To propose and demonstrate an All-Spin Artificial Neural Network (AS-ANN) architecture.
• To utilize a single spintronic device as the fundamental building block for both neuron and synapse functionalities.
• To achieve ultra-low power consumption in neural computing platforms.

Main Methods:

• Development of a novel neural architecture employing a single spintronic device.
• Integration of spintronic devices with CMOS transistors for inter-layer communication.
• Device-level simulations calibrated to experimental results for circuit and system-level analysis.
• Evaluation of the neural network's performance on a standard pattern recognition task.

Main Results:

• The proposed AS-ANN architecture successfully mimics both neuron and synapse functionalities using a single nanoelectronic device.
• Ultra-low voltage operation was achieved due to the low-resistance magneto-metallic neurons, enabling low-voltage operation of spintronic synapses.
• Simulation studies indicated approximately 100x energy savings compared to conventional CMOS implementations.
• The architecture demonstrated effectiveness in a standard pattern recognition problem.

Conclusions:

• The All-Spin Artificial Neural Network represents a significant advancement in neuromorphic computing.
• The single-device approach for neuron and synapse emulation leads to highly power-efficient neural architectures.
• This work paves the way for ultra-low power neural computing platforms leveraging spintronic devices.