The Liquid State Machine (LSM) is a recurrent spiking neural network designed for efficient processing of spatio-temporal streams of information. LSMs have several inbuilt features such as robustness, fast training and inference speed, generalizability, continual learning (no catastrophic forgetting), and energy efficiency. These features make LSM’s an ideal network for deploying intelligence on-device.
In general, single LSMs are unable to solve complex real-world tasks. Recent literature has shown emergence of hierarchical architectures to support temporal information processing over different time scales. However, these approaches do not typically investigate the optimum topology for communication between layers in the hierarchical network, or assume prior knowledge about the target problem and are not generalizable.
In this thesis, a deep Liquid State Machine (deep-LSM) network architecture is proposed. The deep-LSM uses staggered reservoirs to process temporal information on multiple timescales. A key feature of this network is that neural plasticity and attention are embedded in the topology to bolster its performance for complex spatio-temporal tasks. An advantage of the deep-LSM is that it exploits the random projection native to the LSM as well as local plasticity mechanisms to optimize the data transfer between sequential layers. Both random projections and local plasticity mechanisms are ideal for on-device learning due to their low computational complexity and the absence of backpropagating error. The deep-LSM is deployed on a custom learning architecture with memristors to study the feasibility of on-device learning. The performance of the deep-LSM is demonstrated on speech recognition and seizure detection applications.
Library of Congress Subject Headings
Neural networks (Computer science); Machine learning; Memristors; Computer architecture--Design
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Soures, Nicholas M., "Deep Liquid State Machines with Neural Plasticity and On-Device Learning" (2017). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus
Available for download on Saturday, January 12, 2019