A Syncretic Perception Architecture for Bridging Optoelectronic Synaptic Dynamics With In-Sensor Reservoir...

Biological visual systems excel at integrating perception, temporal feature extraction, and information processing with exceptional energy efficiency. Although artificial synapses have demonstrated promising capabilities for integrated sensing and complex computation, their intrinsic nonlinear dynamics remain insufficiently exploited in conjunction with learning algorithms, thereby limiting system-level visual intelligence. Here, we propose a hardware-software co-designed in-sensor reservoir computing platform enabled by a CsPbBr₃/IGZO optoelectronic synapse. The hetero-structured device exhibits persistent photoconductivity, nonlinear relaxation dynamics, and wavelength-dependent synaptic plasticity, providing device dynamics that are useful for reservoir computing. The history-dependent current response translates optical pulse timing into electrical states for temporal encoding in the proposed computing framework. Based on these physical dynamics, a leaky integrate-and-fire neuron-based reservoir computing framework achieves 87% test accuracy on the Fashion-MNIST dataset using five-step spike encoding. Furthermore, a graph-inspired SpikeRC framework is developed to co-optimize device dynamics and spike encoding for color image recognition, achieving 91% accuracy on the CIFAR-10 dataset. This work proposes a hardware-software co-design paradigm that bridges optoelectronic synaptic devices with reservoir intelligence, highlighting a device-to-algorithm approach that couples local optical sensing and temporal encoding with visual recognition.