Data-In-situ Computing with One-Pixel-Multiple-Memristor Architecture for Neuromorphic Sequential Vision

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Traditional neuromorphic vision systems face limitations in dynamic image processing due to inefficient pixel-to-memristor architectures and inadequate temporary storage.
• Existing systems struggle with energy efficiency and latency, hindering real-world applications.

Purpose of the Study:

• To propose a novel one-pixel-multiple-memristor (1PnR) architecture inspired by human visual working memory.
• To develop a data-in-situ computing network for efficient image processing within neuromorphic vision systems.
• To overcome the limitations of traditional architectures in terms of speed and energy consumption.

Main Methods:

• Implemented a one-pixel-multiple-memristor (1PnR) architecture with a rolling exposure strategy for rapid sequential image acquisition.
• Developed a data-in-situ computing network enabling direct computation on stored image data via voltage vectors.
• Eliminated energy-intensive data transmission by performing computations directly on the memristor array.

Main Results:

• Achieved 95.7% recognition accuracy on the Weizmann human action flow dataset with a hardware prototype.
• Demonstrated an estimated 2000× reduction in latency for image sensing and storage compared to CMOS systems.
• Showcased an estimated 160× reduction in energy consumption for image processing.

Conclusions:

• The proposed 1PnR architecture offers a significant advancement in energy efficiency and speed for neuromorphic vision.
• This approach holds substantial potential for developing next-generation, low-power artificial vision systems.
• The data-in-situ computing network effectively addresses data transmission bottlenecks in current systems.