P. Liotet, P. Abry, R. Leonarduzzi, M. Senneret, L. Jaffrès, G. Perrin
The objective of this research is to study the ability of deep learning architectures to learn the temporal dynamics of multivariate time series.
The methodology consists in using well-known synthetic stochastic processes for which the variations in the joint temporal dynamics can be controlled. This allows deep learning to be compared with classical machine learning techniques based on wavelet representations.
First, we evaluate the performance of several different deep learning architectures and show the relevance of convolutional neural networks (CNNs). Second, we test the robustness of CNN performance in classifying subtle changes in multivariate temporal dynamics with respect to training conditions (dataset size, time series sample size, transfer learning).
This paper was published in ICASSP 2020, 45th International Conference on Acoustics, Speech and Signal Processing.