Decoupling Strategy to Separate Training and Inference with Three-Dimensional Neuromorphic Hardware Composed of...

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computer Science

Background:

• Monolithic 3D integration of neuron and synapse devices offers a path towards energy-efficient and compact neuromorphic hardware.
• Current challenges include optimizing performance for both training and inference, which demand distinct synapse device characteristics (endurance vs. retention).

Purpose of the Study:

• To propose and demonstrate a decoupling strategy for training and inference in monolithically integrated 3D neuromorphic hardware.
• To achieve reliable spiking neural network (SNN) operation by utilizing specialized synapse devices for distinct computational phases.

Main Methods:

• Layer-by-layer fabrication of 3D neuromorphic hardware.
• Integration of single-transistor neurons (1T-neurons).
• Incorporation of two distinct synapse types: charge-trap based single thin-film transistor synapses (1TFT-synapses) for inference and memristor synapses (1M-synapses) for training.

Main Results:

• 1TFT-synapses exhibit long retention properties suitable for inference.
• 1M-synapses demonstrate robust endurance for repetitive training updates.
• The hybrid architecture successfully decouples synaptic functions, enabling efficient training and reliable inference.

Conclusions:

• The proposed hybrid synapse architecture effectively addresses the conflicting requirements of training and inference in neuromorphic hardware.
• This decoupling strategy enhances the reliability and performance of SNNs implemented on monolithic 3D integrated systems.
• The layer-by-layer fabrication approach facilitates the integration of diverse synaptic functionalities within a single 3D structure.