As deep neural networks (DNNs) proceed to advance laptop imaginative and prescient, pure language processing, and medical diagnostics, their reliance on 32-bit full-precision weights imposes substantial mannequin dimension and computational burdens that hinder deployment on the edge; to enhance effectivity, we undertake ternary neural networks (TNNs). Right here, we current a ternary circuit composed of two parallel indium–gallium–zinc-oxide (IGZO) thin-film phototransistors (TFPTs) and a resistor, exhibiting three steady, discrete present states ‘OFF’, ‘Intermediate’, and ‘ON’ that map to the ternary weight set {−1, 0, 1}; we additional understand a compact ternary inverter utilizing solely two IGZO TFPTs and two resistors, avoiding advanced binary CMOS logic. The processing path begins with optical sensing, whereby the incident gentle energy densities and wavelength decide discrete voltage outputs; throughout preprocessing, these voltages discretize pixel values (0–255) into a number of intervals which might be provided to the TNN for picture recognition. Leveraging this built-in sensing, preprocessing and inference {hardware} module, we obtain >90% accuracy on CIFAR-10, thereby validating device-level information discretization and transformation and charting a path towards built-in neuromorphic imaginative and prescient methods.
