[AUDITORY] FW: PhD opportunity at Centre for Digital Music (Mark Sandler )

Subject: [AUDITORY] FW: PhD opportunity at Centre for Digital Music From: Mark Sandler <mark.sandler@xxxxxxxx> Date: Tue, 14 Jan 2025 10:24:50 +0000 --_000_DU0PR07MB86195243B811085E2BD782A0BB182DU0PR07MB8619eurp_ Content-Type: text/plain; charset="Windows-1252" Content-Transfer-Encoding: quoted-printable This is an updated posting: the deadline is approaching. Understanding the learning dynamics of Neural Audio models using Linear Alg= ebra Since around 2016 most research in Digital Music and Digital Audio has adop= ted Deep Learning techniques. These have brought important advances in perf= ormance in applications like Music Source Separation, Automatic Music Trans= cription, Timbre Transfer and so on. This is good, but on the downside, the= models get larger, they consume increasingly large amounts of power for tr= aining and inference, require more data and become less understandable and = explainable. These are the issues that underpin the research in this PhD. A fundamental building block in Deep Learning models is Matrix (or Linear) = Algebra. Through training, the matrix that represents each layer is progres= sively modified to reduce the error between a predicted value and the train= ing data. By examining what happens to these matrices during training, it i= s possible to engineer them to learn faster and more efficiently, as well a= s to build DL models that are more compact. Here we turn to Low Rank matric= es: we wish to explore what happens when Low Rank is imposed as a training = constraint in Neural Audio models. Is the model better trained, or not? Is = the model easier and cheaper to train, or not? Early results in non-audio/m= usic applications show that they are better trained and they are cheaper to= train. This work needs developing further for this PhD. Research will start with Music Source Separation, exploring the learning dy= namics of established models like DeMucs. It will then use the knowledge of= these dynamics to intelligently prune the models using the Low Rank approa= ch above [1]. This will speed up the learning and inference and improve pe= rformance. Next, the work could shift to look at other Neural Audio models = and applications or could become more immersed in field of Mechanistic Inte= rpretability, [2], to reveal the hidden, innermost structures that emerge w= ithin trained Neural Networks. Other lines of enquiry could include the tra= de-off between data set size (for training) vs the Ideal Rank of the variou= s layers in the model. Again, early results surprisingly suggest that Low R= ank layers can be trained with less data! Candidates will have excellent background in Linear Algebra (eg Eigenvector= s, Singular Value Decomposition, Tensor Analysis) as well as strong interes= t in some aspect of music or audio. They will also need background in Deep = Learning and a sound knowledge of appropriate programming tools. Knowledge = of Mathematica and the Wolfram Language would be a bonus. You will need a s= trong undergraduate degree and preferably a Masters degree to a high level. Please note that a studentship is only available for those qualifying for C= hina Scholarship Council awards or those qualifying for our faculty=92s S&E= Doctoral Research Studentships for Underrepresented Groups . Self-funded c= andidates are also welcome. Full application guidelines can be found here: https://www.c4dm.eecs.qmul.ac.uk/news/2024-11-12.PhD-call-2025/ For further details of this research topic, contact Mark Sandler (mark.sand= lerl@xxxxxxxx<mailto:mark.sandlerl@xxxxxxxx>) by email. [1] B. Bermeitinger, T. Hrycej, and S. Handschuh, =91Singular Value Decompo= sition and Neural Networks=92, Jun. 2019. doi: 10.1007/978-3-030-30484-3_13= <https://doi.org/10.1007/978-3-030-30484-3_13>. [2] N. Cammarata et al., =91Thread: Circuits=92, Distill, vol. 5, no. 3, p.= e24, Mar. 2020, doi: 10.23915/distill.00024<https://doi.org/10.23915/disti= ll.00024>. [3] V. S. Paul and P. A. Nelson, =91Matrix analysis for fast learning of ne= ural networks with application to the classification of acoustic spectra=92= , The Journal of the Acoustical Society of America, vol. 149, no. 6, pp. 41= 19=964133, Jun. 2021, doi: 10.1121/10.0005126<https://doi.org/10.1121/10.00= 05126>. -- Please note I work part time Monday - Thursday so there may be a delay to m= y email response. professor mark sandler, FREng, CEng, FIEEE, FAES, FIET director of the centre for digital music (c4dm) school of electronic engineering and computer science, queen mary universit= y of london mark.sandler@xxxxxxxx<mailto:mark.sandler@xxxxxxxx> | +44 (0)20 7882 76= 80 --_000_DU0PR07MB86195243B811085E2BD782A0BB182DU0PR07MB8619eurp_ Content-Type: text/html; charset="Windows-1252" Content-Transfer-Encoding: quoted-printable <html xmlns:o=3D"urn:schemas-microsoft-com:office:office" xmlns:w=3D"urn:sc= hemas-microsoft-com:office:word" xmlns:m=3D"http://schemas.microsoft.com/of= fice/2004/12/omml" xmlns=3D"http://www.w3.org/TR/REC-html40"> <head> <meta http-equiv=3D"Content-Type" content=3D"text/html; charset=3DWindows-1= 252"> <meta name=3D"Generator" content=3D"Microsoft Word 15 (filtered medium)"> <style></style> </head> <body lang=3D"en-001" link=3D"#467886" vlink=3D"#96607D" style=3D"word-wrap= :break-word"> <div class=3D"WordSection1"> <div id=3D"mail-editor-reference-message-container"> <div> <div> <div> This is an updated posting: the deadline is approaching= . <o:p></o:p>  <o:p></o:p><= /span> Understanding the learning dynamics of Neural Audio mod= els using Linear Algebra<o:p></= o:p>  <o:p>= </o:p> Since around 2016 most research in Digital Music and Digit= al Audio has adopted Deep Learning techniques. These have brought important= advances in performance in applications like Music Source Separation, Automatic Music Transcription, Timbre Transf= er and so on. This is good, but on the downside, the models get larger, the= y consume increasingly large amounts of power for training and inference, r= equire more data and become less understandable and explainable. These are the issues that underpin the res= earch in this PhD.<o:p></o:p></span= >  <o:p></o:p><= /span> A fundamental building block in Deep Learning models is Ma= trix (or Linear) Algebra. Through training, the matrix that represents each= layer is progressively modified to reduce the error between a predicted value and the training data. By examining wh= at happens to these matrices during training, it is possible to engineer them to learn faster and= more efficiently, as well as to build DL models that are more compact. Her= e we turn to Low Rank matrices: we wish to explore what happens when Low Ra= nk is imposed as a training constraint in Neural Audio models. Is the model better trained, or not? Is the model = easier and cheaper to train, or not? Early results in non-audio/music appli= cations show that they are better trained and they are cheaper to train. This work n= eeds developing further for this PhD.<o:p></o:p>  <o:p></o:p><= /span> Research will start with Music Source Separation, explorin= g the learning dynamics of established models like DeMucs. It will then use= the knowledge of these dynamics to intelligently prune the models using the Low Rank approach above [1].  This will sp= eed up the learning and inference and improve performance. Next, the work c= ould shift to look at other Neural Audio models and applications or could b= ecome more immersed in field of Mechanistic Interpretability, [2], to reveal the hidden, innermost structures that eme= rge within trained Neural Networks. Other lines of enquiry could include th= e trade-off between data set size (for training) vs the Ideal Rank of the v= arious layers in the model. Again, early results surprisingly suggest that Low Rank layers can be trained wit= h less data!<o:p></o:p>  <o:p></o:p><= /span> Candidates will have excellent background in Linear Algebr= a (eg Eigenvectors, Singular Value Decomposition, Tensor Analysis) as well = as strong interest in some aspect of music or audio. They will also need background in Deep Learning and a sound know= ledge of appropriate programming tools. Knowledge of Mathematica and the Wo= lfram Language would be a bonus. You will need a strong undergraduate degre= e and preferably a Masters degree to a high level.<o:p></o:p>=  <o:p></o:p><= /span> Please note that a studentship is only available for th= ose qualifying for China Scholarship Council awards or those qualifying for= our faculty=92s S&= E Doctoral Research Studentships for Underrepresented Groups </= i>= . Self-funded candidates are also welcome.<o:p></o:p>  <o:p></o:p><= /span> Full application guidelines can be found here:<span= style=3D"font-size:12.0pt"><o:p></o:p> &nbsp= ;<a href=3D"https://www.c4dm.eecs.q= mul.ac.uk/news/2024-11-12.PhD-call-2025/" title=3D"https://www.c4dm.eecs.qm= ul.ac.uk/news/2024-11-12.PhD-call-2025/">https://www.c4dm.eecs.qmul.ac.uk/news/2024-11-12.PhD-call-2025= /</a><o:p></o:p>  <o:p></o:p><= /span> For further details of this research topic, contact Mark S= andler (<a href=3D"mailto:mark.sandlerl@xxxxxxxx">mark.sandlerl@xxxxxxxx</a>= ) by email.<o:p></o:p>  <o:p></o:p><= /span> [1] B. Bermeit= inger, T. Hrycej, and S. Handschuh, =91Singular Value Decomposition and Neu= ral Networks=92, Jun. 2019. doi: <a href=3D"https://doi.org/10.1007/= 978-3-030-30484-3_13">10.1007/978-3-030-30484-3_13</a>.<o:p></o:p> [2] N. Cammara= ta et al., =91Thread: Circuits=92, Distill, vol. 5, no. 3, p. e24, Mar. 2020, doi: <a href=3D"https://doi.org/10.23915/distill.00024"><s= pan style=3D"font-size:10.0pt;font-family:Helvetica">10.23915/distill.00024= </a>.<o:p></o:p> [3] V. S. Paul= and P. A. Nelson, =91Matrix analysis for fast learning of neural networks = with application to the classification of acoustic spectra=92, The Journal of the Acoustical Society of America, vol. 149, no. 6, p= p. 4119=964133, Jun. 2021, doi: <a href=3D"https://doi.org/10.1121/= 10.0005126">10.1121/= 10.0005126</a>.<= span style=3D"font-size:12.0pt"><o:p></o:p>  <= span style=3D"font-size:12.0pt"><o:p></o:p>  <o:p></o:p>  <o:p></o:p> <div> <div> -- <o:p></o:p> Please note I work pa= rt time Monday - Thursday so there may be a delay to my email res= ponse. <o:p></o:p><= /p> professor mark sandler, FREng, CEng, FIEEE, FAES, FIET director of the centre for digital music (c4dm) school of electronic engineering and computer science, queen mary= university of london <a href=3D"mailto:mark.sandler@xxxxxxxx= .ac.uk" title=3D"mailto:mark.sandler@xxxxxxxx">mark.sandler@xxxxxxxx</a>&nbs= p;| +44 (0)20 7882 7680 <o:p></o:p>=  <o:p></o:p> </div> </div>  <o:p></o:p></= span> </div> </div> </div> </div> </div> </body> </html> --_000_DU0PR07MB86195243B811085E2BD782A0BB182DU0PR07MB8619eurp_--

This message came from the mail archive
postings/2025/
maintained by:

DAn Ellis <dpwe@ee.columbia.edu>
Electrical Engineering Dept., Columbia University