Distributed Semi-Supervised Learning and Audio Recognition of Road Surfaces

Abstract

The automotive industry is a promising environment for machine learning. However, current machine learning techniques do not meet all the requirements of many possible applications. Requirement such as privacy preservation, limited communication and semi-supervision. To satisfy these requirements, this thesis proposes a simple distributed semi-supervised algorithm (distributed FixMatch). Furthermore, we apply this algorithm to a real-world problem, detecting road surface types from audio. In applying the semi-supervised algorithm to this problem, we also propose a simple augmentation technique for audio features. The proposed algorithm was tested on two real datasets, where the algorithm was compared to a supervised training algorithm. The results suggest that the algorithm successfully leveraged unlabeled data. Furthermore, a theoretical analysis and a simulation show that the communication cost of the proposed algorithm was lower than federated or centralized alternatives.