Highly Tunable Synaptic Modulation in Photo-Activated Remote Charge Trap Memory for Hardware-Based Fault-Tolerant Learning

Abstract

The rapid expansion of deep learning applications for unstructured data analysis has led to a substantial increase in energy consumption. This increase is primarily due to matrix-vector multiplication operations, which dominate the energy usage during inference. Although in-memory computing technologies have alleviated some inefficiencies caused by parallel computing, they still face challenges with broader computational algorithms required for advanced deep learning models. In real-world data collection scenarios, datasets often contain "noisy labels" (errors in annotations), which cause recognition inefficiencies in conventional in-memory computing. Here, a hardware-based fault-tolerant learning algorithm designed for artificial synapses with tunable synaptic operation is proposed. In this scheme, the devices simultaneously process both learning and regulatory signals, enabling selective attenuation of weight updates induced by mistraining signals. Utilizing a high synaptic tunability ratio of 4380 realized in photo-activated remote charge trap memory devices based on defect-engineered hexagonal boron nitride(h-BN), the system nearly completely suppresses weight update signals from mislabeled data, which leads to improved recognition accuracy on the mislabeled Modified National Institute of Standards and Technology (MNIST) dataset. These results demonstrate that tunable synaptic devices can enhance training efficiency in in-memory computing systems for mislabeled datasets, thereby reducing the need for extensive data cleansing and preparation.