.png){kind=icon}
.png){kind=icon}
Return
End-to-end closed-loop optoelectronic computing breaking precision-accuracy coupling
.png){kind=icon}
PRE
AI
DOI：10.1117/1.AP.8.1.016005.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid growth of artificial neural networks has strained conventional electronic processors due to intensive matrix operations and frequent iterative dataflow. Although optoelectronic heterogeneous computing presents a promising pathway through synergistic photonic-electronic integration, its practical implementation remains hindered by disjointed training-inference workflows and offline weight updates. These constraints exacerbate information entropy loss, degrading computational precision and ultimately compromising inference accuracy. Here, we present a phase-pixel lattice-based programmable optical processing unit (OPU) and develop a Lyapunov-based stability theory for its flexible programming. Building upon this, we propose an optoelectronic heterogeneous end-to-end closed-loop architecture (ECA) featuring a programmable OPU tightly combined with electronic processing units. Through end-to-end hardware-algorithm co-design, the ECA network constructs a bidirectional feedback mechanism to compensate for information entropy loss, which fundamentally bridges the training-computing divide. The ECA achieves a precision scaling law by noise-embedded backpropagation, enabling joint photonic-electronic parameter optimization and adaptive computational precision compensation. The inference accuracy using a 4-bit precision OPU reaches 90.8%, matching the 8-bit precision electronic implementation under traditional computing architecture (TCA, 0.1% deviation) and exceeding the current state-of-the-art 4-bit photonic precision under TCA by 1.4% (89.4%). This co-design strategy decouples the precision-accuracy correlation, establishing an architecture for high-performance computing.
Keywords:
optics
photonics
optical computing
artificial neural networks
precision
accuracy