Multivariate Time Series Classification Using Conceptors: Exploring Methods Using Astronomical Object Data

Vlegels, Jamie (2022) Multivariate Time Series Classification Using Conceptors: Exploring Methods Using Astronomical Object Data. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelor's_Project_Thesis_Jamie_Vlegels.pdf Download (435kB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (129kB)

Abstract

Time series classification tasks can be found in a wide range of real-world application domains, ranging from human activity recognition to electrocardiogram diagnostics. Echo state networks have already proven to be excellent tools for such tasks, providing a natural expression of temporal dynamics within their recurrent neural connections. Conceptors, neuro-computational mechanisms, extend the scene of recurrent neural networks by providing high-level conceptual and logical control on the low-level dynamics of a network. During their conception, a demonstration displayed state-of-the-art performance on a time series classification task using a conceptor classification scheme. This paper builds upon this demonstration to further explore conceptors for time series classification. To do this, five variations to the original conceptor classifier were proposed and evaluated on an astronomical object benchmark dataset. Additionally, an analysis of the performance of a heuristic for determining the value of a key conceptor parameter, aperture, was provided. It was found that all of the proposed conceptor classifier variations and baseline echo state network classifiers were able to achieve top-level performance on the benchmark dataset compared to previous methods applied to this dataset. Furthermore, the aperture analysis has shown that a significant increase in conceptor classifier performance can be achieved when choosing the value of the aperture parameter carefully.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Jaeger, H.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	07 Oct 2022 14:33
Last Modified:	07 Oct 2022 14:33
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28813

Actions (login required)

View Item