Learning Rules and Topologies for Liquid State Machines: A Survey of Performance and Representational Dynamics for Image and Speech Recognition

O'Loughlin, Ryan (2022) Learning Rules and Topologies for Liquid State Machines: A Survey of Performance and Representational Dynamics for Image and Speech Recognition. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text OLoughlin_thesis_final_2022.pdf Download (12MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (162kB)

Abstract

As prospect of the next hardware revolution looms, attending to unconventional computing paradigms becomes more important than ever. Recurrent spiking neural networks (RSNNs) are highly suitable for neuromorphic hardware and come ready with an actionable reservoir computing scheme known as the Liquid State Machine (LSM). An LSM employs an RSNN “reservoir” to process input such that meaningful features can be learned with a simple linear readout map trained on reservoir states. LSMs have garnered success in a number of domains, such as speech, image, and even video recognition. However, the hyper-parameter space for reservoir design is prohibitively large and all of these above mentioned successes make use of specialized LSMs, often with significant design-and-tuning overhead. Therefore, a pertinent task for the advancement of LSM literature is the development of an increasingly robust playbook for good reservoir design. To this end, we here survey LSMs in their many forms, both prominent and novel, with particular attention to learning rules and topologies. We characterize reservoir quality as a balance between diverse computational efficacy on two input modalities (image and speech) and coherent dynamics in representational state space. In doing so, we aim to distill favorable reservoir attributes and therefore better equip the state-of-the-art for LSMs that may exhibit the sort of versatility we would hope for in our artificial systems.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Schippers, M.B. and Borst, J.P. and Tiotto, T.F. and Pourcel, G. A. and Abreu, S.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	29 Aug 2022 09:10
Last Modified:	29 Aug 2022 09:10
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28538

Actions (login required)

View Item