PhD Position F/M Incremental Deep Learning for Embedded Systems

PhD Position F/M Incremental Deep Learning for Embedded SystemsLevel of qualifications required :Graduate degree or equivalentOther valued qualifications :Computer Science, Neuroscience, Math, StatisticFonction :PhD PositionAbout the research centre or Inria departmentThe Inria Rennes - Bretagne Atlantique Centre is one of Inria's eight centres and has more than thirty research teams. The Inria Center is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative PMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institute, etc.ContextContextThis PhD will occur in the context of the project Adapting (https://www.pepr-ia.fr/en/projet/adapting-2/) fromthe PEPR AI (https://www.pepr-ia.fr/en/pepr/). This project focuses on designing adaptive embeddedhardware architectures for AI. In this context, our team wants to design new incremental machine learningalgorithms that could serve as use cases in the Adapting project for other researchers who will focus on thehardware architecture design.Continual learning , also known as lifelong learning or incremental learning, is a machine learningparadigm that focuses on the ability of a model to learn and adapt continuously over time as it encountersnew data, tasks, or concepts, without forgetting or catastrophically overwriting what it has previously learned.In continual/incremental learning, the learned model should retain knowledge about previous tasks or datawhile incorporating new information. In this PhD, we will focus on designing new resource-efficientincremental learning algorithms that can run on embedded systems with their associated ressource andprivacy constraints. These contraints involve limited computational power, memory, and energy efficiency.They also involve real-time processing with low latency and often deterministic behavior. Updatingembedded models is complex due to hardware limitations and the need for efficient updates while handlingdata locally to enhance privacy and security.This PhD will focus on foundation models such as well-known LLM -Large Language Models- (e.g. GPT-3.5,Mixtral, Llama 3,...) or multimodal ones (involving for example ViT -Vision Transformer- models such asGPT-4o, Sora, Dall-E 3) and their ability to evolve continuously in an embedded environnement.Application processThe position is funded for 3 years (this is the standard duration of a PhD in France). The net salary is around$2000$ euros. The PhD student will be based in Rennes (https://en.wikipedia.org/wiki/Rennes) and will makea few stays in Grenoble during the 3-years contract.Applications will be processed on a first-come, first-served basis until June 15, 2024.Application Material and ProcedureHere is the supervision team:Martial Mermillod, Professor (in Cognitive Sciences) UGA. MIA chair on "Core AI-Artificial Neural Networks" ([email protected]).Applicants should send these documents to the entire supervision team :An academic CV.An official transcript listing classes taken and grades received.A short statement (maximum of 2 pages) that explains what your research interests are and whyyou would like to work with us.A copy of your most recent English report (e.g. your master report).Any published and relevant research papers (in English, preferably PDF format). If you have severalpapers, please indicate which ones you consider the three most important ones.A list of references (e.g., supervisors)AssignmentRelated workIncremental learning consists in a multi-stage training in which the context (data domain, classes, or task)evolves between each training stage. This paradigm faces the stability-plasticity dilemma. It means that themodel must continuously adapt to new contexts while avoiding catastrophic forgetting (performance on pastcontexts must not deteriorate). The naive approach is the standard fine-tuning strategy in which theparameters of the model (or part of it) are updated from one stage to another by training on the new contextonly. It is known to be efficient for adaptation but it is prone to catastrophic forgetting. Several works focusedon replay strategies to tackle this challenge. It consists in managing a buffer of examples from pastcontexts, which are preserved through the different training stages. Alternatives have beenproposed to avoid storing old data, for instance by learning to preserve the latent representation on the newdata through the training stage . However, current incremental foundation models do not include importantdesiderata for embedded models (e.g. lack of privacy, restricted hardware resources and low energyconsumption, robustness, no need for an oracle or neurogenesis). Our current project will explore the models(1, 2, 3,4,5,6,7) and methods having those important requirements for edge AI. Another line of research is the parameter-oriented approaches: the goal is to associate part of the model's parameters with a specific context in order to maximize the performance on both past and current contexts. This is mainly done by mapping parametersto contexts , parameter pruning or parameter addition . Among the parameter-oriented techniques, arecent trend focused on prompt tuning, used for transformer architectures . It consists in learning inputtokens which are prepended to the input to condition the behavior of the model whose weights are frozen.While parameter pruning and mapping lead to saturation after many training stages, the parameter addictionapproach lead to a growing number of parameters.A complementary approach is the knowledge-editing strategy which aims at correcting specific knowledge by locating and updating the responsible neurons.Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, GregorySlabaugh, Tinne Tuytelaars, "A continual learning survey: Defying forgetting in classification tasks",IEEE Trans. Pattern Anal. Mach. Intell (TPAMI), 44:7, 2022.Lester et al., "The Power of Scale for Parameter-Efficient Prompt Tuning", Conference on EmpiricalMethods in Natural language Processing (EMNLP), 2021 https://arxiv.org/pdf/2104.08691.pdf Meng et al., "Mass-editing Memory in a Transformer", International Conference on Learningrepresentations (ICLR), 2023Main activitiesObjectivesThe objective of this PhD is to investigate incremental learning of foundation models on embedded systems.The graduate student will initially address the question: "Are foundation models prone to catastrophicforgetting with standard fine-tuning techniques?" by conducting comprehensive studies and experiments.Subsequently, she/he will explore how state-of-the-art (SOTA) methods can be applied to these models,assessing the extent to which these existing techniques provide effective solutions to the problem ofincremental learning. This phase will involve adapting and optimizing SOTA methods to suit the constraintsand requirements of embedded systems (efficiency, privacy, low energy consumption, robustness, etc.).Finally, the PhD student will develop and propose new methodologies specifically designed to enablefoundation models to learn incrementally when deployed on embedded systems, ensuring that these modelsmaintain performance and adaptability over time.What we are looking forThe ideal candidate will possess the following skills and attitudes:A master degree in computer science, AI, or a closely related discipline.A strong background in computer science and mathematics.A scientific attitude and the ability to reason through problems.Excellent programming skills.The ability to communicate written and orally in English in a clear and precise manner.A pro-active and independent attitude as well as the ability to function well in a team environment.A good motivation for pursuing a Ph.D. and working in Rennes or Grenoble.Skills(see ideal candidate before)Benefits packagePartial reimbursement of public transport costRemunerationmonthly gross salary amounting to 2100 eurosTheme/Domain :Data and Knowledge Representation and ProcessingStatistics (Big data)(BAP E) Warning : you must enter your e-mail address in order to save your application to Inria. Applications must be submitted online on the Inria website. Processing of applications sent from other channels is not guaranteed.Instruction to applyPlease submit online : your resume, cover letter and letters of recommendation eventuallyDefence Security :This position is likely to be situated in a restricted area (ZRR), as defined in Decree No. 2011-1425 relating to the protection of national scientific and technical potential (PPST).Authorisation to enter an area is granted by the director of the unit, following a favourable Ministerial decision, as defined in the decree of 3 July 2012 relating to the PPST. An unfavourable Ministerial decision in respect of a position situated in a ZRR would result in the cancellation of the appointment.Recruitment Policy :As part of its diversity policy, all Inria positions are accessible to people with disabilities.What we are looking forThe ideal candidate will possess the following skills and attitudes:A master degree in computer science, AI, or a closely related discipline.A strong background in computer science and mathematics.A scientific attitude and the ability to reason through problems.Excellent programming skills.The ability to communicate written and orally in English in a clear and precise manner.A pro-active and independent attitude as well as the ability to function well in a team environment.A good motivation for pursuing a Ph.D. and working in Rennes or Grenoble.About InriaInria is the French national research institute dedicated to digital science and technology. It employs 2,600 people. Its 200 agile project teams, generally run jointly with academic partners, include more than 3,500 scientists and engineers working to meet the challenges of digital technology, often at the interface with other disciplines. The Institute also employs numerous talents in over forty different professions. 900 research support staff contribute to the preparation and development of scientific and entrepreneurial projects that have a worldwide impact. #J-18808-Ljbffr