Enabling Selective and Tunable Weight Updates in All-InGaZnO 3-Transistor 1-Capacitor Synaptic Circuits for On-Chip Training

Abstract

Capacitor-based analog synaptic circuit arrays proposed so far typically required more than three transistors per synapse to enable selective updates for parallel backpropagation updates. For the first time, an innovative update scheme that enables selective updating without requiring additional transistors is demonstrated. This approach is experimentally validated through an all-InGaZnO (IGZO) thin-film transistor (TFT) 3-transistor 1-capacitor (3T1C) synaptic circuit. IGZO TFTs are specifically chosen for their ability to extend retention times due to extremely low leakage currents and their simplified fabrication processes at low temperatures. Fundamental synaptic operations, including controllable weight updates, long data retention, and stable programming endurance, are confirmed experimentally. Additionally, optimizing operational voltage conditions improves weight update behavior, which enhances network training performance. System-level analysis using a neural network hardware simulator with a derived weight update model demonstrates high training accuracy on the MNIST handwritten digit dataset and achieves maximum accuracy over 98%. With the proposed selection method and tunable weight updates, 3T1C synaptic circuit is a promising candidate for scalable large-scale deep neural network accelerators based on analog compute-in-memory technology.