An Energy Efficient FinFET-based Field Programmable Synapse Array (FPSA) Feasible for One-shot Learning on EDGE AI

Abstract

A pure logic 14nm FinFET with capabilities of linearly tunable V-th and excellent retention has been implemented as synapses in neuromorphic system., a Field Programmable Synapse Array (FPSA) has been adopted to replace conventional R-based memory Synapse Array (RSA). Thanks to the wide range of V-t-tuning ability, 200X on/off ratio, and the ultra-small variability, 12%, results showed that the training power and SN ratio of FPSA are 10 times and 50 times smaller than those of the RSA, respectively. Two applications were demonstrated on FPSA array for one-shot learning applications. First, FPSA is used to detect handwritten digits of MNIST dataset. "Learned it by once" can be achieved in this task. Furthermore, FPSA has been applied to recognize goldfish in Cifar 100 dataset after learned the other 4 fish species. With the assistance from one-shot learning, results show the machine learned it faster and better on EDGE. This demonstrates the feasibility of FPSA for low-power and cost-effective synapse-based one-shot learning applications in the AIoT era.