Mixed-signal circuits for embedded machine-learning applications

Machine learning algorithms are attractive solutions for a number of problems in data analytics and sensor signal classification. However, to enable the deployment of such algorithms in embedded hardware, significant progress must be made to reduce the large power dissipation of current GPU and FPGA-based implementations. Our work studies the trade-off between energy and accuracy in neural networks, and looks to incorporate mixed-signal design techniques to achieve low power dissipation in a semi-programmable ASIC implementation.