Evaluating Dimensionality Reduction of 2D Histogram Data from Truck On-board Sensors

Evaldas Vaiciukynas, Matej Ulicny, Sepideh Pashami, Slawomir Nowaczyk

Abstract: This work presents evaluation of several approaches for unsupervised mapping of raw sensor data from Volvo trucks into low-dimensional representation. The overall goal is to extract general features which are suitable for more than one task. Comparison of techniques based on t-distributed stochastic neighbor embedding (t-SNE) and convolutional autoencoders (CAE) is performed in a supervised fashion over 74 different 1-vs-Rest tasks using random forest. Multiple distance metrics for t-SNE and multiple architectures for CAE were considered. The results show that t-SNE is most effective for 2D and 3D, while CAE could be recommended for 10D representations. Fine-tuning the best convolutional architecture improved low-dimensional representation to the point where it slightly outperformed the original data representation.

Conflicts: ktu.lt, hh.se

Keywords: Unsupervised Learning, Deep learning, Applications