Turing AI Fellowship: Adaptive, Robust, and Resilient AI Systems for the FuturE
Lead Research Organisation:
University of Leicester
Department Name: Mathematics
Abstract
The Office for Artificial Intelligence (AI) estimates that AI could add £232 billion to the UK economy by 2030, increasing productivity in some industries by 30%. However, to be truly transformational, the integration of AI throughout the global economy requires understanding and trust in the AI systems deployed. The super-human ability for decision-making in new AI systems requires huge volumes of data with thousands of variables, dependencies and uncertainties. Unregulated application of uncertified data-driven AI, limited by data bias and a lack of transparency, brings huge risks and necessitates a community-wide change. AI systems of the future must also be able to learn on-the-job to avoid becoming a high-interest credit card of huge technical debt. There is thus a timely and unmet need for a new theory and framework to enable the creation and analysis of data-driven AI systems that are adaptive, resilient, robust, explainable, and certifiable, with provable and practically relevant performance guarantees.
This ambitious fellowship, ARaISE, will deliver a radically new framework for the creation of beneficial data-driven AI systems advancing far beyond classical theories by including certifiable robustness and learning in the problem setting. These new theories will enable a formal understanding of the fundamental limits of large-scale data-driven AI, independent of the application area and learning algorithms. This will enable AI practitioners, through understanding such limitations, to influence policy and prevent incidents before they occur.
By connecting different and disparate areas of AI and Machine Learning, working with a world-class team of experts, and by engaging with stakeholders across strategic UK industries and sectors (Healthcare, Manufacturing, Space and Earth Observation, Smart Materials, and Security), ARaISE will create high-value, trustworthy, transformative and responsible AI, capable of reliably 'learning on-the-job' from humans to guarantee capability and trust. Novel human-centric AI, designed to function for the benefit of society, will complement and connect to existing work in the AI research arena, enabling co-development with project partners and focus on strategic industry challenges to ensure real-world relevance is built into research programme and its outputs, facilitating capacity and capability growth. ARaISE will generate gold standard tools for tasks that are currently heavily reliant upon human input and will support long-term global transformation.
Impact and knowledge exchange activities, embedded throughout this programme of work, will support uptake of developed novel AI systems and, through leadership and ambassadorial activities, will support a step-change in how AI systems are built and maintained to ensure resilient, robust, adaptive and trustworthy operation.
The inclusive research programme has been designed to support the career development of the project team and wider stakeholder group maximising the potential for flexible career paths whilst maintaining flexibility to creatively support the team to develop exciting new technology with real world relevance and guide future AI research.
The issues of AI and ethics underpin the programme with responsible research and innovation embedded throughout its activities. Raising public and AI practitioners' awareness, and ultimately influencing policy by active engagement with the UK and AI ethics expertise and policymakers, will ensure that the outcomes are socially beneficial, ethical, trusted and deployable in real world situations. Planned engagement with the ATI, CDTs, partners, and their networks, the development of new partnerships, methodologies and applications, will encourage links between these organisations, build UK expertise, skills and capacity in AI and contribute to realising government investment in UK Societal Challenges and ensure that the UK remains at the forefront of the AI revolution.
This ambitious fellowship, ARaISE, will deliver a radically new framework for the creation of beneficial data-driven AI systems advancing far beyond classical theories by including certifiable robustness and learning in the problem setting. These new theories will enable a formal understanding of the fundamental limits of large-scale data-driven AI, independent of the application area and learning algorithms. This will enable AI practitioners, through understanding such limitations, to influence policy and prevent incidents before they occur.
By connecting different and disparate areas of AI and Machine Learning, working with a world-class team of experts, and by engaging with stakeholders across strategic UK industries and sectors (Healthcare, Manufacturing, Space and Earth Observation, Smart Materials, and Security), ARaISE will create high-value, trustworthy, transformative and responsible AI, capable of reliably 'learning on-the-job' from humans to guarantee capability and trust. Novel human-centric AI, designed to function for the benefit of society, will complement and connect to existing work in the AI research arena, enabling co-development with project partners and focus on strategic industry challenges to ensure real-world relevance is built into research programme and its outputs, facilitating capacity and capability growth. ARaISE will generate gold standard tools for tasks that are currently heavily reliant upon human input and will support long-term global transformation.
Impact and knowledge exchange activities, embedded throughout this programme of work, will support uptake of developed novel AI systems and, through leadership and ambassadorial activities, will support a step-change in how AI systems are built and maintained to ensure resilient, robust, adaptive and trustworthy operation.
The inclusive research programme has been designed to support the career development of the project team and wider stakeholder group maximising the potential for flexible career paths whilst maintaining flexibility to creatively support the team to develop exciting new technology with real world relevance and guide future AI research.
The issues of AI and ethics underpin the programme with responsible research and innovation embedded throughout its activities. Raising public and AI practitioners' awareness, and ultimately influencing policy by active engagement with the UK and AI ethics expertise and policymakers, will ensure that the outcomes are socially beneficial, ethical, trusted and deployable in real world situations. Planned engagement with the ATI, CDTs, partners, and their networks, the development of new partnerships, methodologies and applications, will encourage links between these organisations, build UK expertise, skills and capacity in AI and contribute to realising government investment in UK Societal Challenges and ensure that the UK remains at the forefront of the AI revolution.
Organisations
- University of Leicester (Collaboration, Lead Research Organisation)
- Tangi0 LTD (TG0) (Collaboration)
- Airbus Group (Collaboration)
- Trauma Audit Research Network (TARN) (Collaboration, Project Partner)
- Tangi0 Ltd (Project Partner)
- Synoptix Limited (Project Partner)
- NIHR Leicester Biomedical Research Ctr (Project Partner)
- AstraZeneca (Project Partner)
- CGI Group (UK) (Project Partner)
- NTT DATA Ltd UK (Project Partner)
- BlueSky International Limited (Project Partner)
- University Hospitals of Leicester NHS Trust (Project Partner)
- Toyota Motor Corporation (Project Partner)
- Visual Management Systems Limited (Project Partner)
Publications
Bac J
(2021)
Scikit-Dimension: A Python Package for Intrinsic Dimension Estimation.
in Entropy (Basel, Switzerland)
Barker J
(2022)
Machine learning in sudden cardiac death risk prediction: a systematic review
in EP Europace
Gorban A
(2021)
High-dimensional separability for one- and few-shot learning
Gorban A
(2023)
Stochastic Separation Theorems: How Geometry May Help to Correct AI Errors
in Notices of the American Mathematical Society
Gorban A
(2021)
High-Dimensional Separability for One- And Few-Shot Learning
Gorban AN
(2021)
High-Dimensional Separability for One- and Few-Shot Learning.
in Entropy (Basel, Switzerland)
Gordleeva S.
(2022)
Situation-based memory in spiking neuron-astrocyte network
Grechuk B
(2021)
General stochastic separation theorems with optimal bounds.
in Neural networks : the official journal of the International Neural Network Society
Houston P
(2024)
Efficient High-Order Space-Angle-Energy Polytopic Discontinuous Galerkin Finite Element Methods for Linear Boltzmann Transport
in Journal of Scientific Computing
Related Projects
| Project Reference | Relationship | Related To | Start | End | Award Value |
|---|---|---|---|---|---|
| EP/V025295/1 | 01/01/2021 | 31/01/2022 | £1,463,403 | ||
| EP/V025295/2 | Transfer | EP/V025295/1 | 01/02/2022 | 31/12/2025 | £1,301,722 |
| Description | The project's ambition is the development of next-generation AI systems which are adaptive, stable, and robust by design. Despite the obvious benefits and capabilities brought by new data-driven AI systems, numerous empirical evidence points to examples of their instability, fragility, and rigidity. In this Fellowship we have advanced our understanding of why and how instabilities occur and emerge in data-driven AI systems. We derived new high-level AI instability criteria which could be applied for testing a broad class of classification models. We have also found a completely new class of instabilities - stealth attacks - and developed a theoretical framework for their analysis, assessment, and identification. These latter instabilities are very different from adversarial data perturbations in that they are induced by changes in the AI structure. Remarkably, we have also shown that there are uncountably large classes of problems in which it is very difficult to discriminate between stable and unstable AI from mere accuracy measures. For these problems, stable and unstable AI can be arbitrarily close to each other too. Guided and motivated by this knowledge, we have a created new theoretical and algorithmic framework for dealing with AI instabilities. The framework, in addition to informing and guiding learning algorithms to avoid identified AI vulnerabilities, enables data-driven AI systems to adapt to errors and learn from them. At the moment, we are capable of capturing and characterising major determinants and invariants of such learning and adaptation. Further work is needed to bridge the gap between current theoretical understanding and practice. The latter problem is closely related to the problem of learning from low-sample high-dimensional data which is one of the fundamental challenges in modern post-classical machine learning, data science, and AI. The other major objectives of the project are the advocacy of AI technologies and the removal of AI adoption barriers. We are achieving this through a combination of engagement activities with the demonstration of appropriate AI design practices ensuring the provable and certifiable robustness of AI systems. These demonstrations are being achieved within use-cases co-developed and co-designed with our industrial partners and other stakeholders. |
| Exploitation Route | The project lays out the foundations and the methodology to design stable, robust, adaptive data-driven AI systems with provable performance guarantees. The methodology is published in open-access outlets and exemplified, where appropriate, in open-source software. Since the project is ongoing, we are still working on completing the theory and the framework as well as working on producing usable and convenient open-source code. |
| Sectors | Aerospace Defence and Marine Digital/Communication/Information Technologies (including Software) Electronics Environment Financial Services and Management Consultancy Healthcare Manufacturing including Industrial Biotechology Security and Diplomacy |
| Description | Despite the project only approaching its mid-term point, the fellowship's research is already having a positive impact on addressing real-life challenges faced by my project partners. In our work with Synoptix Ltd we implemented various AI error correction approaches to reduce the ratio of false positives in their products (with one of such products being currently going through live trials at several rail level crossings in the UK). With Tangi0, we used the kernel versions of stochastic separation theorems in the customisation of the proprietary gesture recognition algorithms. With TARN, we contributed to the creation of new predictive models for the quantitative assessment of outcomes in patients after major trauma. We hope that the models will be used nationwide. There were some unexpected impacts, achieved through the collaboration with the Arch-I-Scan project (funded by the AHRC), were we have developed a novel way to overcome the challenge of insufficient data and increase the accuracy of the classification of archaeological artefacts. The method has been demonstrated as a proof of concept, and we are working on a joint publication with the Arch-I-Scan team. To date, the Fellowship has advanced the understanding of the new concentration of measure phenomenon - stochastic separation theorems - and is set to establish a new theoretical framework of continuous learning with provable guarantees. It has discovered a new class of AI vulnerabilities - stealth attacks - and developed a basis for the theoretical and computational assessment of these vulnerabilities. We have also developed a novel framework to understand instabilities in AI. All these developments seeded several new lines of inquiries which we are planning to explore in the coming years. These include the development of new formulations of the learning problem which could explicitly account for the non-stationarity of environments and the conservativism of classical distribution-agnostic settings. Another example is the creation of new evolutionary learning algorithms guided by new knowledge discovered in the Fellowship: the importance and relevance of various measures of intrinsic dimension to learning. Through the Fellowship's support of academic engagement, we are exploring the creation of new theories and practical implementations of in-material computation, going beyond classical von Neumann computations. |
| First Year Of Impact | 2022 |
| Sector | Electronics,Healthcare,Culture, Heritage, Museums and Collections,Transport |
| Impact Types | Cultural Societal Economic Policy & public services |
| Description | Data: a new direction: A call for evidence from Department for Digital, Culture, Media & Sport |
| Geographic Reach | National |
| Policy Influence Type | Contribution to a national consultation/review |
| Description | Gave evidence at the All Party Parliamentary Group on AI evidence meeting concerning national security: Regulation of AI-driven live facial recognition technologies |
| Geographic Reach | National |
| Policy Influence Type | Participation in a guidance/advisory committee |
| Description | Trauma Audit Research Network |
| Geographic Reach | National |
| Policy Influence Type | Contribution to new or improved professional practice |
| Description | 101158046 - AUTOMATA |
| Amount | € 4,887,940 (EUR) |
| Funding ID | 101158046 - AUTOMATA |
| Organisation | European Commission |
| Sector | Public |
| Country | Belgium |
| Start | 08/2024 |
| End | 02/2029 |
| Description | AI technologies for the next generation of quantum imaging |
| Amount | £78,500 (GBP) |
| Organisation | Royal Commission for the Exhibition of 1851 |
| Sector | Charity/Non Profit |
| Country | United Kingdom |
| Start | 09/2021 |
| End | 09/2024 |
| Description | AUTOMated enriched digitisation of Archaeological liThics and cerAmics |
| Amount | € 4,887,940 (EUR) |
| Funding ID | Project: 101158046 - AUTOMATA |
| Organisation | European Commission H2020 |
| Sector | Public |
| Country | Belgium |
| Start | 08/2024 |
| End | 03/2029 |
| Description | EP/V025295/1 additional GRIP funding |
| Amount | £17,500 (GBP) |
| Organisation | Engineering and Physical Sciences Research Council (EPSRC) |
| Sector | Public |
| Country | United Kingdom |
| Start | 09/2021 |
| End | 04/2022 |
| Description | ESA AO/1-10605/21/NL/KML -Towards a Thermal Digital Twin |
| Amount | £55,250 (GBP) |
| Organisation | European Space Agency |
| Sector | Public |
| Country | France |
| Start | 08/2021 |
| End | 02/2023 |
| Description | Impact of the COVID-19 Pandemic on Risk Factor Control and Outcomes in Patients |
| Amount | £227,763 (GBP) |
| Organisation | Health Data Research UK |
| Sector | Charity/Non Profit |
| Country | United Kingdom |
| Start | 12/2021 |
| End | 09/2022 |
| Description | Industry Fellowship |
| Amount | £98,000 (GBP) |
| Organisation | Royal Academy of Engineering |
| Sector | Charity/Non Profit |
| Country | United Kingdom |
| Start | 08/2023 |
| End | 09/2025 |
| Description | Self-learning AI-based digital twins for accelerating clinical care in respiratory emergency admissions (SLAIDER) |
| Amount | £619,667 (GBP) |
| Funding ID | EP/Y018281/1 |
| Organisation | Engineering and Physical Sciences Research Council (EPSRC) |
| Sector | Public |
| Country | United Kingdom |
| Start | 09/2023 |
| End | 04/2025 |
| Title | AI error correctors |
| Description | The tool enables and automates creation of AI error correctors |
| Type Of Material | Improvements to research infrastructure |
| Year Produced | 2022 |
| Provided To Others? | Yes |
| Impact | Exact impact is not known yet, however the tool has already been successfully used in several other ongoing projects. |
| URL | https://github.com/Mirkes/Error-corrector |
| Title | MATLAB implementation of stealth attacks |
| Description | The tool enables the analysis of vulnerabilities of AI structures to stealth attacks. |
| Type Of Material | Improvements to research infrastructure |
| Year Produced | 2022 |
| Provided To Others? | Yes |
| Impact | The tool enables the identification of vulnerabilities in neural networks to a new class of security threats - stealth attacks. These threats have been determined and analised in our work: Tyukin, I.Y., Higham, D.J., Bastounis, A., Woldegeorgis, E. and Gorban, A.N., 2021. The feasibility and inevitability of stealth attacks. arXiv preprint arXiv:2106.13997. |
| URL | https://github.com/tyukin/Stealth-adversarial-attacks |
| Title | Python package Scikit-dimension (2021) |
| Description | We have contributed to the development of a Python package for computing various measures of data intrinsic dimension. The package can be accessed at https://github.com/j-bac/scikit-dimension. Description of the package is published in a technical note in Entropy: Bac, J., Mirkes, E.M., Gorban, A.N., Tyukin, I. and Zinovyev, A., 2021. Scikit-dimension: a python package for intrinsic dimension estimation. Entropy, 23(10), p.1368. |
| Type Of Material | Improvements to research infrastructure |
| Year Produced | 2021 |
| Provided To Others? | Yes |
| Impact | This is a comprehensive package enabling computation of various measures of data intrinsic dimension. The computation is relevant in Neural Architecture Search (as is evidenced in our work Zhou, Q., Gorban, A.N., Mirkes, E.M., Bac, J., Zinovyev, A. and Tyukin, I.Y., 2022, July. Quasi-orthogonality and intrinsic dimensions as measures of learning and generalisation. In 2022 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE) and also for the analysis and understanding of how the information flows through deep learning structures/networks. In view of the relationships between dimensionality, stability, learning, and generalisation (see our research outputs), the tool is instrumental to address these important fundamental questions. |
| URL | https://github.com/j-bac/scikit-dimension |
| Title | Tensorflow-based framework for construction of modified Concept Cells architecture |
| Description | A framework for quick high-level construction of Concept Cells models for classification problems. Framework developed upon previous basic Concept Cells examples (MATLAB source) and extended to Tensorflow for higher impact upon open source. |
| Type Of Material | Improvements to research infrastructure |
| Year Produced | 2023 |
| Provided To Others? | No |
| Impact | Quick construction of the architecture and GPU-accelerated training. The TensorFlow-based framework allows more general community access and utilisation. |
| Title | MyI-Net: Fully Automatic Detection and Quantification of Myocardial Infarction from Cardiovascular MRI Images |
| Description | Myocardial infarction (MI) occurs when an artery supplying blood to the heart is abruptly occluded. The "gold standard" method for imaging MI is cardiovascular magnetic resonance imaging (MRI) with intravenously administered gadolinium-based contrast (with damaged areas apparent as late gadolinium enhancement [LGE]). However, no "gold standard" fully automated method for the quantification of MI exists. We have developed a new end-to-end fully automatic system (MyI-Net) for the detection and quantification of MI in MRI images. The model is described in Wang, S., Abdelaty, A.M., Parke, K., Arnold, J.R., McCann, G.P. and Tyukin, I.Y., 2023. MyI-Net: Fully Automatic Detection and Quantification of Myocardial Infarction from Cardiovascular MRI Images. Entropy, 25(3), p.431. |
| Type Of Material | Computer model/algorithm |
| Year Produced | 2023 |
| Provided To Others? | No |
| Impact | The model forms the basis for automated quantification of MI. Working jointly with BRC Leicester, are developing the model further with the view to testing the model in clinical environments. |
| Description | ESA Airbus Digital Thermal Twin |
| Organisation | Airbus Group |
| Department | Airbus Defence and Space UK |
| Country | United Kingdom |
| Sector | Private |
| PI Contribution | The work has just recently started, as a part of our ESA AO/1-10605/21/NL/KML -Towards a Thermal Digital Twin project. The University is a subcontractor in this task, with several Schools engaged. Our part is to contribute to the development of an AI module capable of correlating mathematical models of a thermal shield to data. |
| Collaborator Contribution | Airbus will provide us with models to correlate as well as empirical data |
| Impact | No outputs emerged yet apart from the award of the grant ESA AO/1-10605/21/NL/KML -Towards a Thermal Digital Twin and initial scoping work |
| Start Year | 2021 |
| Description | Innovate UK KTP012250 between TG0 Limited and University of Leicester (Artificial Intelligence technology for learning and recognition of tactile gestures) |
| Organisation | Tangi0 Ltd (TG0) |
| Country | United Kingdom |
| Sector | Private |
| PI Contribution | In this partnership, which I initiated with Tangi0 whilst preparing my Fellowship application, we focused on the development of a light-weight AI module for automated gesture recognition using smart materials. In addition to contributing into the development of the module, the team contributed by training Tangi0 staff involved in the project. The project, whilst planned in 2020, started in 2021. |
| Collaborator Contribution | The partner is contributing by providing access to hardware and technology and by providing pathways to impact. |
| Impact | The project is still active and we are envisioning economic, societal, and policy & public services impacts occurring as a result of this work |
| Start Year | 2020 |
| Description | Quantification of scar volumes from MRI - BRC Leicester |
| Organisation | University of Leicester |
| Department | NIHR Biomedical Research Centre |
| Country | United Kingdom |
| Sector | Hospitals |
| PI Contribution | Analysed data, developed models for scar segmentation and build error correctors to filter out false positives |
| Collaborator Contribution | Images from 600 cases |
| Impact | A manuscript has now been submitted and published. A joint grant application prepared Wang, S., Abdelaty, A.M., Parke, K., Arnold, J.R., McCann, G.P. and Tyukin, I.Y., 2023. MyI-Net: Fully Automatic Detection and Quantification of Myocardial Infarction from Cardiovascular MRI Images. Entropy, 25(3), p.431. |
| Start Year | 2021 |
| Description | TARN - prediction of patients' outcomes after major trauma |
| Organisation | Trauma Audit Research Network (TARN) |
| Country | United Kingdom |
| Sector | Academic/University |
| PI Contribution | Analysed TARN data, developed AI/ML models, and explored the possibility for developing AI correctors for these models |
| Collaborator Contribution | Provided cleaned and processed data |
| Impact | No outcomes emerged yet - a manuscript is currently being prepared |
| Start Year | 2021 |
| Description | AI for East Midlands Universities (December 2022) |
| Form Of Engagement Activity | A formal working group, expert panel or dialogue |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Professional Practitioners |
| Results and Impact | The activity aimed at bringing together expertise from East Midlands Universities around AI. One of the outcomes was the formation of a working collaborative partnership between my group at KCL and the research group from the University of Loughborough working on neuromorphic and memristive networks and AI models. |
| Year(s) Of Engagement Activity | 2022 |
| Description | AIDAM Winter AI Workshop (December 2022) |
| 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 | Gave a lecture on the origins of AI errors and the mathematical framework and tools we are developing to address these. The audience responded with great interests as was evidenced by in-depth discussions afterwards. |
| Year(s) Of Engagement Activity | 2022 |
| Description | Advances in Bio-inspired and Neuromorphic Electronics Workshop 2022 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Delivered an invited talk at the event aimed at the UK community working on bio-inspired and neuromorphic computations. The audience represented relevant research centres across the UK including the University of Edinburgh, University College Cork, Oxford, University of Sheffield, KCL, University of Bristol, University of the West of Scotland, UCL, Liverpool, Texas A&M University, University of California Berkley, Technical University of Dresden, Max Plank institute, Salk Institute, University of Southern California |
| Year(s) Of Engagement Activity | 2022 |
| Description | EAI International Conference on IoT and Big Data Technologies for HealthCare |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Keynote lecture presented: I.Tyukin. The challenge of trustworthy data- driven Artificial Intelligence: boundaries, limitations, and opportunities, October 18-19, 2021 |
| Year(s) Of Engagement Activity | 2021 |
| Description | Encouraging Enterprise at KCL (March 2023) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Local |
| Primary Audience | Professional Practitioners |
| Results and Impact | The goal of the event was to encourage and advocate engagement of King's academic communities with industries and other stakeholders in the area of AI. The lecture was very well received by the audience who appreciated the advice. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Few-shot learning (Sheffield, July 2022) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | Approximately 30 people attended the talk which has led to the discussion around the foundational problem of learning from few examples in high-dimensional and low-dimensional settings. A collaborative activity around the analysis of the notion of dimension of data sampled from trajectories of dynamical systems. The project involves academics from the University of Sheffield, Leicester, and King's College London. We have also agreed to work on a grant proposal and have already submitted one outline proposal (AI Hubs: Computational and Mathematical Foundations) |
| Year(s) Of Engagement Activity | 2022 |
| Description | ICERM Safety and Security of Deep Learning |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | A talk on the subject of mathematical quantification and high-level tests of non-symbolic AI / data-driven AI vulnerabilities I.Tyukin. Breaking into a Deep Learning box, April 10 - 11, 2021. ICERM Safety and Security of Deep Learning |
| Year(s) Of Engagement Activity | 2021 |
| Description | International Joint Conference on Neural Networks 2021 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | A talk given explaining why and when few-shot learning works: Ivan Y. Tyukin, Alexander N. Gorban, Muhammad H. Alkhudaydi, Qinghua Zhou. Demystification of Few-shot and One-shot Learning. International Joint Conference on Neural Networks, 2021. |
| Year(s) Of Engagement Activity | 2021 |
| Description | International Joint Conference on Neural Networks 2022 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | A presentation at the World Congress on Computational Intelligence / International Joint Conference on Neural Networks |
| Year(s) Of Engagement Activity | 2022 |
| Description | Invited seminar at University College London and Imperial College London |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | Invited to give a seminar jointly to members of University College London and Imperial College London. Interesting discussions followed about extending our work into a completely different field. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Invited seminar at University College London and Imperial College London |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | Invited to give a seminar jointly to members of University College London and Imperial College London. Demonstration of impact of theoretical advances on practical research and applications of AI systems. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Invited seminar on the limitations of data-driven AI, the hierarchy of AI instabilities, and approaches to address them (Cambridge 2023) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | About 15 people attended the seminar followed by a discussion of further methods and tools needed for better understanding of the phenomenon. An agreement has been reached to work on several related topics and on the application of the tools to the analysis of dynamical systems. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Invited seminar on the limitations, verifiability, and the challenge of AI errors (Newcastle 2023) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | Approximately 20 people attended the seminar where I presented our latest findings on the inevitable typicality of AI errors. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Invited talk at Loughborough C-Dice (Centre for postdoctoral development in infrastructure, cities, and industry) 2023 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Postgraduate students |
| Results and Impact | The talk aimed at the development of new-generation talent in AI and Cognitive Technologies. I shared my personal experience and lessons learned on the pathway to becoming an independent academic. I also shared a personal reflection on how to build for successful proposals and, importantly, reflected on the spectrum of responsibilities attached to winning a large grant. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Invited talk at the China University of Mining and Technology (November 2022) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Presented our tools, stochastic separation theorems, to the broader international community of early career researchers and academics. The presentation instigated a very interesting debate about the feasibility of learning in large-scale systems, its limitations, and applications to real-life problems. |
| Year(s) Of Engagement Activity | 2022 |
| Description | Isaac Newton Institute workshop: Mathematics of Deep Learning |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | A lecture was delivered on the subject of security and vulnerabilities in data-driven AI models I.Tyukin. Breaking into Deep Learning models, October 5, 2021, Isaac Newton Institute. |
| Year(s) Of Engagement Activity | 2021 |
| Description | Keynote talk at SOFSEM 2025 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Keynote talk at SOFSEM 2025. This is a traditional computer science meeting with a very long history, attended by leading experts in theoretical computer science |
| Year(s) Of Engagement Activity | 2025 |
| URL | http://www.sofsem.sk/ |
| Description | Learning and generalisation from low-sample high-dimensional data (Edinburgh, October 2022) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Regional |
| Primary Audience | Professional Practitioners |
| Results and Impact | Approximately 60 people attended the seminar, which was followed by the discussion around the need to rethink the problem of learning in the context of the prevalence of distribution-agnostic approaches. New collaboration with Alex Serb and his research group emerged as a result of the activity. This was followed by a hackathon with two PDRAs participating from our side and one PDRA from Edinburgh. We are now working towards a joint paper summarising the findings. |
| Year(s) Of Engagement Activity | 2022 |
| Description | Learning from Few Examples with Nonlinear Feature Maps. London Computing Conference 2023 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | The work attracted major attention and was awarded the Best Paper Award at the Conference. |
| Year(s) Of Engagement Activity | 2023 |
| Description | NeurIPS 2024 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Presentation at one of the most prestigious conferences in the area of Artificial Intelligence. The work revealed new vulnerabilities of Large Language Models and proposed a novel method of editing Large Language Models without damaging their existing skills and with provable guarantees |
| Year(s) Of Engagement Activity | 2024 |
| URL | https://neurips.cc/Conferences/2024 |
| Description | Neuromorphic tuning of feature spaces to overcome the challenge of low-sample high-dimensional data, IJCNN 2023 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | This was a talk at a major international conference on neural networks. The presentation facilitated several questions and a discussion afterwards. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Presenting at IEEE IJCNN 2024 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | I gave a talk at IEEE IJCNN 2024 |
| Year(s) Of Engagement Activity | 2024 |
| URL | https://2024.ieeewcci.org/ |
| Description | Stochastic Separation Theorems for making AI Safe, Adaptive, and Robust. International Congress on Industrial and Applied Mathematics, 2023. |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | The talk raised interesting debate on the importance of the development of mathematical methods and tools to characterise robustness and stability of AI systems |
| Year(s) Of Engagement Activity | 2023 |
| Description | THE INEVITABILITY OF AI ERRORS AND THE CHALLENGE OF TRAINING ACCURATE AND VERIFIABLY STABLE DATA-DRIVEN AI. A lecture at the Eindhoven Artificial Intelligence Systems Institute |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | The talk raised an intense debate around the notions of stability in AI and led to a joint proposal submitted to the University of Eindhoven for a PhD project in this area. We are also considering developing a joint grant application with colleagues from Eindhoven. |
| Year(s) Of Engagement Activity | 2023 |
| Description | Talk and research visit at the Norwegian University of Science and Technology (NTNU) |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Went on a research visit to collaborators at the Norwegian University of Science and Technology (NTNU). Gave a seminar on recent research, attended by approximately 30 people, from senior academincs, to masters' students. Outlined new possible directions for collaborative research. |
| Year(s) Of Engagement Activity | 2022 |
| Description | Talk at AIDAM workshop, University of Leicester |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | Local |
| Primary Audience | Professional Practitioners |
| Results and Impact | Many students and ECR attended, the talk sparked a fruitful discussion with professionals and practitioners working in the area of AI in medicine and healthcare |
| Year(s) Of Engagement Activity | 2024 |
| Description | Talk at International Conference on Scientific Computation and Differential Equations (SciCADE) 2022 |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Invited to present recent research findings at an international conference. Talk was attended by approximately 80 people. Discussions about the work later led to a new collaboration with colleagues at the Norwegian University of Science and Technology (NTNU). |
| Year(s) Of Engagement Activity | 2022 |
| URL | https://scicade2021.hi.is/ |
| Description | Workshop on Complex and simple models of multidimensional data : from graphs to neural networks |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | The purpose of the workshop was to create an international forum for a broad discussion around the topics of complexity, dimensionality, and artificial intelligence and disseminate state-of-the-art knowledge in the area. The event was held online, with more than 150 registrations. In addition to organising the workshop a lecture on the topic of certifiable learning from low-sample high-dimensional data was delivered too. I.Tyukin, A.N. Gorban. The mathematics of learning from high-dimensional low-sample data with small neural networks, December 1, 2021. Workshop on Complex and simple models of multidimensional data : from graphs to neural networks. |
| Year(s) Of Engagement Activity | 2021 |
| Description | Workshop on Robust, Stable, and Secure AI (November 2022) |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | I organised the workshop and invited world-leading experts to participate. The workshop brought together experts across both academia and industry to debate the challenge of AI errors and the impact these errors have. All participants confirmed the importance of the topic and reported on the new perspectives and directions the workshop brought into their own research. |
| Year(s) Of Engagement Activity | 2022 |
| Description | Workshop on Robustness and Stability of Neural Networks at the 2023 International Conference on Artificial Neural Networks, Crete, September 2023 |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | The workshop attracted significant attention of the delegates. Members of my groups and myself received invitations to visit other international groups and present our work there |
| Year(s) Of Engagement Activity | 2023 |