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

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Materials Science

Background:

• Deep learning's growth increases energy demands, mainly from matrix-vector operations during inference.
• In-memory computing offers efficiency but struggles with complex algorithms and noisy labels in real-world data.
• Current systems face recognition inefficiencies due to errors in data annotations (noisy labels).

Purpose of the Study:

• To propose a hardware-based, fault-tolerant learning algorithm for artificial synapses with tunable operations.
• To enable selective attenuation of weight updates caused by mistraining signals.
• To enhance training efficiency and recognition accuracy in in-memory computing systems with noisy datasets.

Main Methods:

• Developed a fault-tolerant learning algorithm for tunable artificial synapses.
• Integrated learning and regulatory signals for selective weight update control.
• Utilized photo-activated remote charge trap memory devices with defect-engineered hexagonal boron nitride (h-BN).

Main Results:

• Achieved a high synaptic tunability ratio of 4380 in h-BN based devices.
• The system effectively suppressed weight update signals from mislabeled data.
• Demonstrated improved recognition accuracy on the mislabeled Modified National Institute of Standards and Technology (MNIST) dataset.

Conclusions:

• Tunable synaptic devices enhance in-memory computing efficiency for datasets with noisy labels.
• The proposed algorithm reduces the need for extensive data cleansing and preparation.
• This approach offers a pathway to more energy-efficient and robust deep learning systems.