A Trainable Synapse Circuit Using a Time-Domain Digital-to-Analog Converter.

We propose a CMOS synapse circuit using a time-domain digital-to-analog converter (TDAC) for realizing spiking neural network hardware with on-chip learning. A TDAC has the advantages that (i) it can reproduce a post-synaptic potential and (ii) its number of analog components for DA conversion, such as current sources and capacitors, is independent of the bit width. We designed a synapse circuit using TSMC 40 nm technology, and the synaptic weight was updated by the remote supervised method (ReSuMe) or spike timing-dependent plasticity (STDP). The circuit simulation results of the designed circuit show that it can execute ReSuMe and generate the time-window function for STDP with high energy efficiency. [ABSTRACT FROM AUTHOR]