Sparse Bayesian reconstruction for optimal facial recognition
Lead Research Organisation:
University of Cambridge
Department Name: Physics
Abstract
Near 100% accuracy is a necessary condition for facial recognition to become more widely utilised in security and to be more widely accepted by users. Existing techniques in the field of computer vision have taken us to unprecedented levels of accuracy. This project aims to close the final gap in accuracy using cutting-edge image analysis techniques developed by astronomers.
Traditional facial recognition software acts directly on a computer processed image to determine an identity. Using a direct image as input has bonuses such as simplicity, but drawbacks in that many theoretical and practical questions are difficult to pose on a pixel-based representation.
In a more established image analysis practice, one traditionally acts on a representation of the image that is more specific to the object of study, in this case the human face. In this compressed representation it becomes easier to ask more specific questions.
This project seeks to apply the cutting edge of such sparse Bayesian representations to the field of facial recognition, using the state-of-the-art in inference algorithms to determine the optimal and minimal unbiased representation of a human face. This project will be able to provide an orthogonal approach to currently applied techniques and could be the missing link in closing the final gap toward 100% accuracy.
Traditional facial recognition software acts directly on a computer processed image to determine an identity. Using a direct image as input has bonuses such as simplicity, but drawbacks in that many theoretical and practical questions are difficult to pose on a pixel-based representation.
In a more established image analysis practice, one traditionally acts on a representation of the image that is more specific to the object of study, in this case the human face. In this compressed representation it becomes easier to ask more specific questions.
This project seeks to apply the cutting edge of such sparse Bayesian representations to the field of facial recognition, using the state-of-the-art in inference algorithms to determine the optimal and minimal unbiased representation of a human face. This project will be able to provide an orthogonal approach to currently applied techniques and could be the missing link in closing the final gap toward 100% accuracy.
People |
ORCID iD |
Anthony Lasenby (Principal Investigator) |
Publications
Alsing J
(2021)
Nested sampling with any prior you like
Alsing J
(2021)
Nested sampling with any prior you like
in Monthly Notices of the Royal Astronomical Society: Letters
Alsing Justin
(2021)
Nested sampling with any prior you like
in arXiv e-prints
Carragher Ethan
(2021)
Convergent Bayesian Global Fits of 4D Composite Higgs Models
in arXiv e-prints
Fowlie A
(2020)
Nested sampling with plateaus
Fowlie A
(2021)
Nested sampling with plateaus
in Monthly Notices of the Royal Astronomical Society
Fowlie A
(2020)
Nested sampling cross-checks using order statistics
in Monthly Notices of the Royal Astronomical Society
Javid Kamran
(2020)
Compromise-free Bayesian neural networks
in arXiv e-prints
Joachimi B
(2021)
When tension is just a fluctuation How noisy data affect model comparison
in Astronomy & Astrophysics
Roque I
(2021)
Bayesian noise wave calibration for 21-cm global experiments
in Monthly Notices of the Royal Astronomical Society
Description | Royal Society University Research Fellowship: Bayesian machine learning and tensions in cosmology |
Amount | £722,622 (GBP) |
Organisation | The Royal Society |
Sector | Charity/Non Profit |
Country | United Kingdom |
Start | 09/2020 |
End | 09/2028 |
Description | AnyVision Innovation Partnership Scheme (IPS) partnership |
Organisation | AnyVision |
Country | Israel |
Sector | Private |
PI Contribution | Ongoing research into problems of interest to the company (assessing bias in facial recognition datasets, training Bayesian Neural Networks as autoencoders) |
Collaborator Contribution | From the collaboration support letter: We envisage AnyVision committing the activity of three research scientists alongside key members of our senior technical team to review and potentially implement the first stage results of Generative Head. This is a real-world commitment to the project from AnyVision of over £250,000. Should the project generate valuable Intellectual Property, PolyChord and AnyVision are discussing mutual ownership of that IP, in order that AnyVision can use its existing routes to market to exclusively commercially exploit the improvements and use them to possibly open new market areas. |
Impact | Project ongoing. |
Start Year | 2020 |
Description | PolyChord Ltd Innovation Partnership Scheme (IPS) partnership |
Organisation | PolyChord Ltd |
Country | United Kingdom |
Sector | Private |
PI Contribution | Ongoing research using their Bayesian inference tools and expertise including access to their latest Bayesian Neural Network training tools and data scientists |
Collaborator Contribution | From partner support letter: PolyChord Ltd will be committing additional contribution to this project in these amounts: Time spent by Data Scientists and members of the management team over 18 months, this will be at least 33 percent of a senior data scientist's time £60K Costs of office, computers and equipment £18K Costs of travel £1.2K |
Impact | Research is ongoing, but use of their tools by our research has spurred further innovation in the company |
Start Year | 2020 |
Description | Mathematical Challenges in the Electromagnetic Environment Workshop, DAMTP, Cambridge UK |
Form Of Engagement Activity | Participation in an activity, workshop or similar |
Part Of Official Scheme? | No |
Geographic Reach | National |
Primary Audience | Professional Practitioners |
Results and Impact | Workshop organised by PA Consulting and DSTL to bring cutting edge mathematics to apply to problems relevant to the defense sector. Will Handley was invited to bring expertise and to give a talk. Academic mathematical experts drawn from across the country, with other policymakers and DSTL members present |
Year(s) Of Engagement Activity | 2020,2021 |
URL | https://gateway.newton.ac.uk/event/tgmw74 |
Description | Nested Sampling: an efficient and robust Bayesian inference tool for physics and machine learning, Physics Colloquium, Adelaide, Australia. |
Form Of Engagement Activity | A talk or presentation |
Part Of Official Scheme? | No |
Geographic Reach | International |
Primary Audience | Professional Practitioners |
Results and Impact | Colloquium talk to a broad audience of particle physicists and astrophysicists |
Year(s) Of Engagement Activity | 2020 |
URL | https://github.com/williamjameshandley/talks/tree/adelaide_2020 |
Description | Nested sampling: an efficient and robust Bayesian inference tool for astrophysics and cosmology |
Form Of Engagement Activity | A talk or presentation |
Part Of Official Scheme? | No |
Geographic Reach | National |
Primary Audience | Professional Practitioners |
Results and Impact | Seminar given to the Oxford University Astrophysics group |
Year(s) Of Engagement Activity | 2020 |
URL | https://github.com/williamjameshandley/talks/tree/oxford_2020 |