Single neuromorphic memristor closely emulates multiple synaptic mechanisms for energy efficient neural networks

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Artificial Intelligence

Background:

• Biological synapses possess complex functions beyond memory and weight multiplication, including short-term plasticity and meta-plasticity.
• Artificial neural networks (ANNs) typically model only long-term memory and weight multiplication.
• Emulating complex synaptic functions in hardware is crucial for advancing neuromorphic computing.

Purpose of the Study:

• To demonstrate memristive nano-devices capable of emulating multiple biological synaptic functions.
• To integrate these multi-functional memristors into bio-inspired deep neural networks (DNNs).
• To evaluate the performance and energy efficiency of these DNNs in a reinforcement learning task.

Main Methods:

• Fabrication of strontium titanate (SrTiO3)-based memristive nano-devices operating in a non-filamentary, low conductance regime.
• Development of bio-inspired DNNs incorporating memristors that emulate long-term memory, short-term memory, short-term plasticity, and meta-plasticity.
• Training the DNN to play the Atari Pong video game, a dynamic reinforcement learning task.

Main Results:

• The memristive devices successfully emulated all targeted complex synaptic functions.
• The bio-inspired DNN demonstrated stable and energy-efficient operation.
• Energy consumption was reduced by approximately two orders of magnitude compared to a GPU implementation for the same task.

Conclusions:

• Memristive devices can effectively emulate complex synaptic functionalities, paving the way for advanced neuromorphic hardware.
• These multi-functional hardware synapses significantly improve the energy efficiency of DNNs.
• This research broadens the applicability of neuromorphic computing and enhances AI performance and energy costs.