School of Electronics & Computer Science

Location:  Highfield Campus
Closing Date:  Friday 25 June 2021
Reference:  1366721FP

Deep Scattering Networks for Efficient Audio Analysis

Supervisory Team:    Dr. Kate Farrahi, Dr. Jagmohan Chauhan, Dr. Stefan Bleeck

Project description

Wavelet scattering networks compute translation invariant representations that are stable to deformations, while preserving high-frequency information. Scattering networks have been developed as a mathematical and algorithmic perspective to addressing the desirable properties that deep convolutional networks can provide but using a cascade of wavelet transforms with nonlinear modulus and averaging operations. These networks have major computational advantages since the operations are fixed (hardcoded in their structure) and therefore do not contain any trainable parameters. In addition to computational savings, scattering networks have the advantage of being more interpretable. 

Scattering networks have been evaluated in the computer vision domain and have shown promising results. They have also been used for image reconstruction and for image generation. However, scattering networks have not been applied to many other applications as they are quite a new topic. This project will focus on the investigation of scattering networks for audio analysis. The goal is to investigate how scattering networks can perform in various prediction tasks related to audio while reducing computation and increasing interpretability. We will investigate the feasibility of scattering networks on a wide range of audio tasks including speech enhancement, scene understanding and sound localization. Depending on the results, we further aim to improve the performance of scattering networks in the audio domain by developing pertaining optimization techniques. We hope that the work done in this project will enhance the capabilities of existing deep learning models being deployed on devices such as smart voice assistants and hearables amongst others.

This project is funded through the UKRI MINDS Centre for Doctoral Training (www.mindscdt.ai). This is one of 16 PhD training centres in the UK with a unique focus on advancing AI techniques in the context of real-world engineered systems with a remit that spans novel hardware for AI, AI and machine learning, pervasive systems and IoT, and human-AI collaboration. We provide enhanced cross-disciplinary training in electronics and AI, entrepreneurship, responsible research and innovation, communication strategies, outreach and impact development as part of an integrated 4-year iPhD programme. 

The MINDS CDT is based in a dedicated laboratory on Highfield Campus at the University of Southampton. The lab provides a supportive environment for individual research, ideas sharing and collaboration, and the wider campus provides access to substantial high-performance computing (including dedicated GPU servers), maker and cleanroom facilities. You will take part in our annual, student-designed innovation camps, be able to work with industry and government partners through our internship scheme and be able to take part in exchanges with international university partners.

Entry Requirements

A very good undergraduate degree (at least a UK 2:1 honours degree, or its international equivalent).

Closing date: 25 June 2021

Funding: full tuition for UK Students an enhanced stipend £18,285, tax-free per annum for 4 years.

How To Apply

Applications should be made online. Select programme type (Research), 2021/22, Faculty of Physical Sciences and Engineering, next page select “iPhD Machine Intelligence for Nano-electronic Devices and Systems. (Full time)”. In Section 2 of the application form you should insert the project title and name of the supervisor. 

Applications should include: 

• Research Proposal
• Curriculum Vitae
• Two reference letters
• Degree Transcripts to date

Apply online: https://www.southampton.ac.uk/courses/how-to-apply/postgraduate-applications.page

For further information please contact: feps-pgr-apply@soton.ac.uk