High-order dynamics in an ultra-adaptive neuromorphic vision device

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Current neuromorphic hardware for artificial general vision intelligence faces limitations in area and power efficiency due to reliance on multiple components.
• A key challenge is the lack of a single device integrating diverse neuronal dynamics across optical and electrical domains.

Purpose of the Study:

• To report a single ultra-adaptive neuromorphic vision device capable of mimicking biological visual processing.
• To demonstrate a device that amalgamates spiking and graded neuronal dynamics with ultralow power consumption.

Main Methods:

• Development of a single ultra-adaptive neuromorphic vision device (I_xT_yO_1-x-y/CuO_x/Pd) with tailored electronic properties.
• Utilizing in situ scanning transmission electron microscopy and technology computer-aided design simulations to elucidate real-time optoelectronic dynamics.
• Construction of an artificial general vision intelligence system using homogeneous device arrays.

Main Results:

• The device integrates broadband retinal spiking neuron and non-spiking graded neuron, and cortical synapse and neuron dynamics.
• Achieved superior power efficiency of up to 67.89 trillion operations per second per watt.
• Demonstrated area efficiency of up to 3.96 mega operations per second per feature size (MOPS/F²).

Conclusions:

• The developed single neuromorphic vision device offers a highly efficient and adaptable solution for artificial general vision intelligence.
• The device supports both event-driven and frame-driven paradigms, enabling versatile cognitive imaging.
• This breakthrough paves the way for next-generation, bio-inspired visual processing systems.