A Brain-Inspired Homeostatic Neuron Based on Phase-Change Memories for Efficient Neuromorphic Computing.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Materials Science

Background:

• Synaptic scaling is a biological homeostatic mechanism crucial for neural network learning.
• Existing hardware spiking neural networks (SNNs) lack intrinsic dynamic regulation mechanisms.
• Phase Change Memory (PCM) devices offer potential for novel artificial neuron designs.

Purpose of the Study:

• To develop a novel artificial neuron using PCM devices for neuromorphic computing.
• To implement homeostatic and plastic phenomena for internal regulation in SNNs.
• To enhance learning capabilities, robustness, and energy efficiency in neuromorphic systems.

Main Methods:

• Designed and fabricated novel PCM-based artificial neurons.
• Integrated these neurons into spiking neural networks (SNNs).
• Evaluated performance in multi-pattern learning, continual learning (CNNs), and autonomous navigation tasks.

Main Results:

• Demonstrated increased robustness and optimized multi-pattern learning under spike-timing-dependent plasticity (STDP).
• Improved continual learning resilience and accuracy in hybrid supervised-unsupervised convolutional neural networks (CNNs).
• Showcased energy efficiency gains via bio-plausible active forgetting enabled by PCM conductance drift.

Conclusions:

• PCM-based neurons enable self-regulating, dynamic SNNs with enhanced learning and robustness.
• These neurons facilitate bio-inspired recurrent networks for autonomous decision-making.
• PCM device characteristics, like conductance drift, can be leveraged for energy-efficient neuromorphic computing.