A Deep Reinforcement Learning Approach to Audio-based Navigation and Audio Source Localization

by Prof. Aggelos Pikrakis (Department of Informatics, University of Piraeus)

Tuesday 12 Apr 2022, 16:00 → 17:00 Europe/Athens

This talk presents a formulation of the problems of environment navigation and source localization from a deep reinforcement learning perspective, as an alternative to more traditional methods, for the case when the only available information is the raw sound coming from the environment. To this end, we discuss how a deep neural network architecture can blend with an online reinforcement learning algorithm so that an autonomous agent can be trained to solve these two problems. We also make an attempt to present certain issues related to the flow of error signals in neural networks during the training stage using the presented method as a starting point. On the application side, we show that the proposed agent architecture achieves good performance and exhibits satisfactory generalization capabilities for both problems user study, even when a limited amount of training data is available or when the environmental parameters change  in ways not encountered during the training stage. Finally, we explain that a certain degree of transfer learning is feasible between these two applications. 

 

Videoconference via https://us02web.zoom.us/j/81907627101

Aggelos Pikrakis_12_04_2022.mp4