Brain-Inspired In-Memory Data Pruning and Computing with TaO_x Mem-Selectors.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Conventional edge vision systems suffer from high power consumption and latency due to separate pruning, memory, and processing units.
• Human visual selective attention offers a model for efficient information processing in edge devices.

Purpose of the Study:

• To develop a multifunctional device for in-memory pruning-computing (IMPC) inspired by human visual attention.
• To investigate the switching mechanisms within a novel Mem-Selector (M-S) device.

Main Methods:

• Fabrication and characterization of a Ta/TaOx/Ta2O5 Mem-Selector device.
• Utilizing transmission electron microscopy (TEM) to observe filament formation and nanocrystalline cluster growth.
• Construction and evaluation of an IMPC system for adaptive information pruning and processing.

Main Results:

• The M-S device exhibits both resistive memory and threshold switching characteristics, indicating coexisting ionic and electronic mechanisms.
• The IMPC system adaptively prunes trivial information, optimizing hardware cost and classification performance.
• Significant reductions in input energy consumption (29-90%) with minimal accuracy loss (<1%) were achieved.

Conclusions:

• The developed M-S device and IMPC system offer a promising solution for energy-efficient, high-performance edge hardware.
• Hardware-software co-design is crucial for advancing edge computing capabilities.
• The study highlights the potential of bio-inspired computing for next-generation AI systems.