Neuromemristive Extreme Learning Machines for Pattern Classification

This paper presents a novel neuromemristive architecture for pattern classification based on extreme learning machines (ELMs). Specifically, we propose CMOS current-mode neuron circuits, memristor-based bipolar synapse circuits, and a stochastic, hardware-friendly training approach based on the least-mean-squares (LMS) learning algorithm. These components are integrated into a current-mode ELM architecture. We show that the current-mode design is especially efficient for implementing constant network weights between the ELM´s input and hidden layers. The neuromemristive ELM was simulated in the Cadence AMS design environment. We used an experimental memristor model based on experimental data from an HfO_{x} device. The top-level design was validated by training a 10 hidden-node network to detect edges in binary patterns. Results indicate that the proposed architecture and learning approach are able to yield 100% classification accuracy.