A Memristor-Based Liquid State Machine for Auditory Signal Recognition

Date of Award


Degree Name

M.S. in Electrical and Computer Engineering


Department of Electrical and Computer Engineering


Christopher G. Yakopcic


Spiking Neural Networks (SNNs) are the third generation of neural networks that incorporate the notion of time in their model. SNNs are starting to be deployed in neuromorphic systems that implement on-device learning with low size, weight, and power (SWaP). This work is specifically interested in audio recognition applications. The temporal nature of audio input streams makes SNNs a natural choice for this task. The Liquid State Machine (LSM) is a special type of recurrent neural network that uses spiking neurons. The inherent features of the LSM, such as robustness, fast training, and energy efficiency, make it a promising alternative to other networks for the processing of complex, spatio-temporal input streams. Instead of training many layers of a neural network, the LSM harnesses the properties of a randomly generated recurrent neural network with fixed connectivity through a dynamic mapping of the input data to a higher-dimensional representation. This enhanced representation of the input information can be converted to linearly separable states, thus avoiding the need to resort to multi-layer, gradient-based backpropagation.The memristor is a non-volatile, nanoscale device that is an attractive candidate for representing the synapses of a neural network with very lower power and area density. This work presents a novel memristor-based LSM architecture capable of classifying different musical styles. To drive the analog spiking neuron circuits with the column outputs of a memristive dot-product circuit, an input voltage converter is simulated in 180nm technology. Also, to store the state vectors used for training, a digital counting circuit is realized in 180nm technology, and a sampling technique is introduced to reduce the overall hardware overhead required. To further prove the feasibility of a neuromemristive, mixed-signal implementation of the LSM system, a spike mapping technique is proposed that can successfully match 90% of the neuron outputs in the hardware and software neuron models. The performance of the memristor-based LSM is demonstrated on a musical style classification experiment, performing 90% accuracy on a two-genre classification task compared to 91% accuracy in software. Furthermore, memristor device non-idealities and faulty neurons were investigated and the proposed system was proven to be resilient to noise and fault-tolerant


Electrical Engineering, Engineering, memristor, liquid state machine, spiking neural network

Rights Statement

Copyright 2021, author