Wearable in-sensor reservoir computing using optoelectronic polymers with through-space charge-transport...

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Artificial Intelligence

Background:

• Biological vision systems exhibit efficient in-sensor multi-task learning, a goal for artificial general intelligence.
• Conventional silicon-based vision chips and deep learning models face challenges in time/energy efficiency, scalability, and affordability for edge devices.

Purpose of the Study:

• To develop a material-algorithm co-design for an affordable and efficient in-sensor learning system.
• To emulate the human retina and an affordable learning paradigm using novel organic materials.

Main Methods:

• Fabrication of a wearable transistor-based dynamic in-sensor Reservoir Computing (RC) system using bottle-brush-shaped semiconducting p-NDI.
• Utilizing efficient exciton dissociation and through-space charge transport characteristics of p-NDI.
• Integration with memristive organic diodes for a 'readout function' to process information.

Main Results:

• The developed RC system demonstrated excellent separability, fading memory, and echo state properties.
• High accuracies were achieved in recognizing handwritten letters (98.04%), numbers (88.18%), and classifying costumes (91.76%).
• The system successfully classified 3 types of hand gestures from event-based videos with 98.62% accuracy, outperforming existing organic semiconductors.

Conclusions:

• The proposed material-algorithm co-design offers a promising pathway for affordable and highly efficient photonic neuromorphic systems.
• This approach significantly reduces computing costs compared to conventional artificial neural networks.
• The developed system advances in-sensor learning capabilities for edge AI applications.