Authors:
Lisa-Maria Beneke
1
;
Michell Boerger
1
;
Philipp Lämmel
1
;
Helene Knof
1
;
Andrei Aleksandrov
1
and
Nikolay Tcholtchev
1
;
2
Affiliations:
1
Fraunhofer Institute for Open Communication Systems (FOKUS), Berlin, Germany
;
2
RheinMain University of Applied Sciences, Wiesbaden, Germany
Keyword(s):
Liquid Time-Constant Neural Networks, LTC, RNN, LSTM, PTB-XL, Time Series Analysis.
Abstract:
Neural networks have become pivotal in timeseries classification due to their ability to capture complex temporal relationships. This paper presents an evaluation of Liquid Time-Constant Neural Networks (LTCs), a novel approach inspired by recurrent neural networks (RNNs) that introduces a unique mechanism to adaptively manage temporal dynamics through time-constant parameters. Specifically, we explore the applicability and effectiveness of LTC in the context of classifying myocardial infarctions in electrocardiogram data by benchmarking the performance of LTCs against RNN and LSTM models utilzing the PTB-XL dataset. Moreover, our study focuses on analyzing the impact of various pre-processing methods, including baseline wander removal, Fourier transformation, Butterworth filtering, and a novel x-scaling method, on the efficacy of these models. The findings provide insights into the strengths and limitations of LTCs, enhancing the understanding of their applicability in multivariate
time series classification tasks.
(More)