Precise deep neural network computation on imprecise low-power analog hardware

There is an urgent need for compact, fast, and power-efficient hardware implementations of state-of-the-art artificial intelligence. Here we propose a power-efficient approach for real-time inference, in which deep neural networks (DNNs) are implemented through low-power analog circuits. Although analog implementations can be extremely compact, they have been largely supplanted by digital designs, partly because of device mismatch effects due to fabrication. We propose a framework that exploits the power of Deep Learning to compensate for this mismatch by incorporating the measured variations of the devices as constraints in the DNN training process. This eliminates the use of mismatch minimization strategies such as the use of very large transistors, and allows circuit complexity and power-consumption to be reduced to a minimum. Our results, based on large-scale simulations as well as a prototype VLSI chip implementation indicate at least a 3-fold improvement of processing efficiency over current digital implementations.