Supervised Learning Using Spike-Timing-Dependent Plasticity of Memristive Synapses.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence Hardware
• Materials Science

Background:

• Spiking neural networks (SNNs) mimic biological brains but face challenges in hardware implementation for efficient learning.
• Error backpropagation is crucial for training deep learning models but difficult to implement directly in SNN hardware.
• Ferroelectric memristors offer promising analog synaptic weight storage for neuromorphic computing.

Purpose of the Study:

• To propose a supervised learning model enabling error backpropagation for spiking neural network hardware.
• To demonstrate the feasibility of this model using a circuit with three-terminal ferroelectric memristors (3T-FeMEMs).
• To evaluate the potential for high-speed and low-power computation in large-scale SNNs.

Main Methods:

• Developed a modified supervised learning model tailored for hardware implementation.
• Designed a network circuit utilizing 3T-FeMEMs as analog synaptic weight storage devices.
• Implemented error backpropagation by reflecting network error to the write voltage of 3T-FeMEMs and incorporating spike-timing-dependent learning.

Main Results:

• Successfully demonstrated the XOR problem using numerical simulations with estimated circuit properties.
• Validated the model's ability to perform benchmark learning tasks.
• Showcased the potential for efficient learning in SNN hardware.

Conclusions:

• The proposed supervised learning model with 3T-FeMEMs enables effective error backpropagation in SNN hardware.
• The learning time per step is independent of the number of neurons, indicating scalability.
• This approach promises high-speed, low-power computation for large-scale artificial intelligence applications.