Opening the 'Black Box' of Neural Networks
Lead Research Organisation:
University of Cambridge
Abstract
Abstracts are not currently available in GtR for all funded research. This is normally because the abstract was not required at the time of proposal submission, but may be because it included sensitive information such as personal details.
Organisations
People |
ORCID iD |
Publications
| Description | 1) We have extended existing techniques used to automatically configure machine learning algorithms, making them applicable to a broader, more useful class of configuration settings. This will benefit a wide variety of applications, improving the final performance obtained from machine learning systems and reducing the time required to tune parameters to achieve that performance. Our work thus far will be published in a paper at the 2022 International Conference on Learning Representations, which proposes several avenues for advancing our results; we hope to investigate several of these in the remainder of the funded period. 2) In follow-up work, we investigated theoretical failings of a broad class of optimisation methods, arguing for an alternative approach which avoided these concerns. Surprisingly, our proposed alternative did not outperform established methods to the degree expected, which has incited further work into the mechanisms underlying existing techniques. |
| Exploitation Route | 1) Our technique is applicable to a wide range of machine learning applications, so will be relevant to practitioners across a range of sectors as ML-driven 'intelligence' becomes more commonly used. We have detailed our approach in a submission to an open-access conference, and have made our research code publicly available, so all interested parties may deploy the logic in their own applications, and other researchers may pick up the ideas to advance further. 2) The unexpected results of our work will form part of my PhD thesis, and we hope to publish them in an open-access venue for further consideration by the research community. |
| Sectors | Digital/Communication/Information Technologies (including Software) |
| Description | Contributed Presentation to 19th EUROPT Workshop on Advances in Continuous Optimization |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Postgraduate students |
| Results and Impact | Contributed a presentation to the "Multiobjective optimization and machine learning" track of the conference, engaging with other researchers with more theoretical and mathematical backgrounds in optimisation. |
| Year(s) Of Engagement Activity | 2022 |
| URL | https://sites.fct.unl.pt/europt2022/home |
| Description | Invited Talk at Vector Institute |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Postgraduate students |
| Results and Impact | Invited to give a talk to the Vector Institute at the University of Toronto, Canada, by a PhD student there whose work was a key foundation in developing our contribution. Attendance included current PhD students and principal investigators. In follow-up question-and-answer, we discussed the outstanding issues in this area of research. |
| Year(s) Of Engagement Activity | 2022 |