An Online Learning Method Using Spike-Timing Dependent Plasticity for Neuromorphic Systems.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Artificial neural networks (ANNs) commonly use gradient descent for training.
• Implementing ANNs on hardware requires efficient, low-power learning mechanisms.
• Spike-timing dependent plasticity (STDP) is a biologically inspired learning rule.

Purpose of the Study:

• To propose and validate an online learning method for ANNs based on STDP.
• To demonstrate hardware implementation of STDP for neural network training.
• To compare the efficacy of STDP-based learning with gradient descent.

Main Methods:

• Simulated a single-layer neural network on a cross-point array using MATLAB for gradient descent training.
• Developed and applied a novel pulse scheme based on STDP to train the same network.
• Utilized a previously reported 4-terminal synaptic transistor model.
• Applied teaching pulses with controlled timing differences to the synaptic transistors' back gate.

Main Results:

• The STDP-based learning method successfully trained binary MNIST samples.
• Synaptic weight maps generated by the STDP method closely matched those from gradient descent.
• The proposed method achieved hardware-based training without reliance on computer calculations.

Conclusions:

• The STDP-based online learning method is a viable hardware implementation for ANN training.
• This approach offers a power-efficient alternative to traditional gradient descent for neuromorphic systems.
• The study validates the potential of STDP for efficient, on-chip learning in synaptic transistor networks.