QMUL School of Electronic Engineering and Computer Science
Centre for Digital Music / Machine Listening Lab Seminar Series
**Seminar by Konstantinos Drossos (Tampere University, Finland) **
**Date/time: 3pm, Wednesday, 6th March 2019 **
**Location: Engineering Building, room 2.16 ** Campus map: https://www.qmul.ac.uk/media/qmul/docs/about/Mile-End_map-May2018.pdf
Open to students, staff, alumni, public; all welcome. Admission is FREE, no pre-booking required.
Title: How it can adapt and what it learned. Domain adaptation and interpretability for deep learning based machine listening methods
Abstract: In a typical scenario we use a dataset to optimize and develop our deep learning based machine listening method, and then we try to deploy it in a realistic environment. Usually there is a drop in the performance, especially when the deployment environment is quite different from the one that our used dataset came from. A promising way of tackling this problem is the domain adaptation, with recent published methods focusing on general audio to provide optimistic results. Although the domain adaptation seem to be an efficient way to make our methods to adapt in new conditions, we still are left with the question of what our method learned during training. This presentation will focus on the problems of domain adaptation and interpretability of deep neural networks, considering the case of general audio and machine listening. For the domain adaptation part, the presentation will include recent results with the underlying theoretical framework for acoustic scene unsupervised domain adaptation. For the interpretability part, the presentation will show some results from a work in progress for understanding what is learned by a typical deep neural networks based method used in the task of sound event detection.
Bio: Konstantinos Drossos is postdoc at the Audio Research Group, Tampere University of Technology, under Prof. Tuomas Virtanen. He has been a postdoc fellow at Montreal Institute for Learning Algorithms (MILA), under Prof. Yoshua Bengio. He holds a BEng in Sound Technology, a BSc in Informatics, an MSc in Sound and Vibration Research, and a PhD in emotional information retrieval from sound events. His main research foci are audio captioning, domain adaptation, interpretability, and machine listening.