Coupled Ferroelectric-Photonic Memory in a Retinomorphic Hardware for In-Sensor Computing.

Area of Science:

• Materials Science
• Nanotechnology
• Artificial Intelligence

Background:

• The Internet-of-Things (IoT) drives demand for energy-efficient, real-time processing in smart sensors.
• In-sensor computing, placing computation at the data source, is crucial for edge processing.
• Existing artificial visual systems seek to emulate biological vision for enhanced performance.

Purpose of the Study:

• To develop a prototype visual sensor integrating photosensitivity and ferroelectricity.
• To mimic human retinal functions like photoreception and memory computation.
• To demonstrate in-sensor image processing capabilities using novel materials.

Main Methods:

• Utilized two-dimensional (2D) α-In2Se3 material with integrated ferroelectricity and photosensitivity.
• Engineered a device mimicking retinal photoreceptors and amacrine cells using electrical switching polarization.
• Implemented in-sensor convolution image processing via a phototransistor network with pre-programmed kernels.

Main Results:

• Achieved ≈94% accuracy in encoding and classifying 12,000 MNIST images using photon-induced short-term plasticity.
• Demonstrated gate-tunable excitatory and inhibitory functions for image analysis.
• Successfully performed in-sensor convolution for edge and feature enhancement in images.

Conclusions:

• Ferroelectric α-In2Se3 shows significant potential for compact and efficient retinomorphic hardware.
• The developed device offers a promising platform for advanced artificial visual systems.
• This approach advances in-sensor computing for real-time visual data processing.