QMUL, School of Electronic Engineering and Computer Science
Centre for Digital Music Seminar Series
Seminar by:
Vipul Arora
Date/time: Wednesday 11th of October, 14:00-15:00
Location: Hybrid
Bancroft Road Teaching Room 4.01 in the Mile End campus, or join zoom meeting.
Open to students, staff, alumni, public; all welcome. Admission is FREE, no pre-booking required.
Title: Model Adaptation for Learning from Small Data
Bio: Vipul Arora is an Associate Professor at the Department of Electrical Engineering, IIT Kanpur, India. He received his B.Tech. and PhD degrees in Electrical Engineering from IIT Kanpur. He did a postdoc at Oxford University (UK), where he developed speech recognition systems using linguistic principles. Then he worked at Amazon in Boston (USA), where he worked on audio classification for developing the Alexa home security system. His current research interest is in developing machine learning algorithms for audio and signal processing. He works on model adaptation, uncertainty modelling and generative machine learning. http://home.iitk.ac.in/~vipular/
Abstract: Deep-learning-based models achieve remarkable performances with big labelled data. However, many practical scenarios face a scarcity of labelled data, while there may still be an abundance of unlabelled data. This talk will discuss several methods to learn effectively from small data. These methods mostly fall under the paradigm of model adaptation and include fine-tuning-based transfer learning, meta-learning, and semi-supervised domain adaptation. These methods’ application to music melody estimation and sensor calibration (regression) problems will be demonstrated. Another way to learn from limited data is by using conditional models. This method will be illustrated for generative machine learning applied to XY models in statistical Physics and field theories in Particle physics. Time permitting, there will be a brief presentation on audio retrieval.
Video recording: https://www.youtube.com/watch?v=-_AS8_NNtWw