UCLA

Artificial neural networks (ANN) have demonstrated performance beyond human capability in challenging games like Go and chess. However, they are still limited by the von Neumann bottleneck, which requires massive overhead for data transmissions between logic and memory units. The end of Moore's Law calls for novel approaches to meet the accelerating computational demands of big data and machine learning. Neuromorphic circuits are promising candidates, inspired by the speed, parallelism, and efficiency of the human brain. A new synaptic device, the synstor, uses Schottky barriers at its I/O terminals and charge trap memory to combine the synaptic functions of convolutional signal processing, Hebbian learning, and nonvolatile analog memory. Unlike circuits based on other synaptic devices, such as memristors and phase change memory, the synstor circuit can spontaneously program its synaptic weights (conductances) via Hebb’s rule, without external computational circuits or complex unit cells. A 20 � 20 crossbar array of synstors is fabricated and connected to artificial neurons to form a self-programming neuromorphic circuit. By applying equal amplitude voltage pulses to synstor input and output electrodes during inference and learning, both processes can run concurrently in a synstor circuit. A 4 � 2 synstor crossbar with 2 “neurons” performs speech recognition with energy efficiency of ~1017 FLOPS/W, surpassing existing computing circuits. Additionally, a 2 � 2 synstor circuit programs itself in real-time to stabilize a morphing wing by modulating its camber in a dynamic wind speed environment. The synstor circuit demonstrates performance superior to human participants and a computer-based controller in this task after repeated trials. If scaled up, synstor circuits have the potential to bypass the von Neumann bottleneck of transistor computing circuits, leading to a new computing platform with real-time self-programming and intelligence in complex dynamic environments.