An Energy Efficient Time-Mode Analog Neural Network

In this work we present an energy efficient mixed-signal realization of a quantized neural network. In this implementation we represent our signals in the time domain. By encoding our signals as pulse-widths, we can take advantage of the extremely accurate clocks and delay resolution in nanometer CMOS processes. Weights are represented as binary weighted current sources with local DRAM storage. Capacitive integration implements weighted addition, and an inverter provides as rectified linear activation. We present the design in 0.18um CMOS process with successful circuit and behavioral simulation results on the IRIS dataset.