Timing-Dependent Spiking Neural Network: Board-Level Hardware Implementation with Photoelectroactive Van der Waals...

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence Hardware
• Materials Science

Background:

• The increasing volume of unstructured data necessitates energy-efficient, real-time computing solutions.
• Biological brains offer a model for efficient computation through spike-timing-dependent plasticity (STDP).

Purpose of the Study:

• To experimentally realize a multi-channel timing-dependent spiking neural network (TD-SNN) at the board level.
• To emulate STDP using novel photoelectroactive synaptic devices integrated with neuron circuits.
• To demonstrate adaptive and real-time learning capabilities in neuromorphic hardware.

Main Methods:

• Integration of photoelectroactive synaptic devices with analog leaky integrate-and-fire (LIF) neuron circuits.
• Emulation of STDP by exploiting timing between electrical presynaptic and optical postsynaptic spikes.
• Engineering presynaptic pulse shapes to achieve diverse STDP learning rules (Hebbian, anti-Hebbian, all-LTP, all-LTD).

Main Results:

• Demonstrated reversible and bidirectional modulation of synaptic weights via photoelectroactive doping.
• Achieved self-learning, system-level adaptation, and competitive behaviors in integrated networks.
• Attained up to 90.9% accuracy on MNIST tasks with STDP parameters, showing robust pattern classification when the LTP/LTD area ratio (PDR) ≥ 1.25.

Conclusions:

• This work presents the first experimental realization of a multi-channel TD-SNN at the board level.
• The developed photoelectroactive synaptic devices enable diverse STDP learning rules and adaptive network behaviors.
• The findings represent a significant advancement in timing-dependent neuromorphic hardware, demonstrating feasibility for adaptive, real-time learning systems.