Back to search

FRINATEK-Fri prosj.st. mat.,naturv.,tek

Unsupervised Lifelong Learning

Alternative title: Ikke-veiledet livslang maskinlæring

Awarded: NOK 12.0 mill.

The overall goal of this research is to establish a novel foundation 'ULEARN: Unsupervised Lifelong Learning' that will shift the future design of machine learning to tackle significant engineering challenges of long running systems operating in ever-changing environments. ULEARN aims to replace the actual manual system design with a dynamic software architecture that implements well-composed learning processes, which are dynamically optimized and evolving in a holistic lifetime manner. The research is centered around finding structure in data, recognizing anomalies during operation, adapting the learning architecture and (re-)organizing and evolving own learning processes at runtime against unforeseen situations. Furthermore, the whole approach will be conceptualized into a novel self-learning system to realize the integrated vision of unsupervised lifelong learning. In the previous project period, the research further developed foundational elements around the challenge of “catastrophic forgetting” in self-supervised lifelong learning. A publication was made at the ICCV workshop on continual learning to contribute with task-agnostic approach that reduces the effects of the “catastrophic forgetting problem”. In this previous period, we also developed a concept on lifelong clustering method to support the unsupervised lifelong learning. The resulted paper was published at ICML. Finally, we also developed several methods on how to deploy unsupervised lifelong learning concept by elaborating on a set of applied methods for general-purpose computer vision systems, federated deep learning and federated convolution transformer that resulted in 6 publications in IEEE transactions and conference proceedings. In the next period, we will focus on further robustifying the unsupervised task-agnostic approach that includes the lifelong clustering for further reducing the catastrophic forgetting and deploying it into general-purpose computer vision systems. Finally, we would like to mention that the international collaboration has been reallocated with University of Pennsylvania (UPENN) as our main collaborator (Rene Vidal) and his team have moved from JHU to UPENN.

ULEARN targets a novel foundation “Unsupervised Lifelong Learning” that will shift the future design of machine learning to tackle significant engineering challenges of long running systems operating in ever-changing environments. Today, the design of learning systems primarily occurs before runtime, requiring continuous updates to tackle changes. In the future hyper-connected digital world, learning systems will need to work radically different than today, ubiquitously connected to a tremendous amount of diverse data and facing unanticipated and a priori unknown conditions during operation. The essence of ULEARN is to replace the actual manual system design with a dynamic software architecture that implements well-composed learning processes, which are dynamically optimized and evolving in a holistic lifetime manner. Computing systems will self-learn, adapting and evolving to the highly dynamic conditions at runtime. Such new designed systems can learn from high-dimensional complex data and are capable to autonomously evolve their own learning processes under changing and unforeseen conditions, lifelong. This new paradigm will be achieved by the concerted research of three PIs, bringing together diverse expertise in unsupervised learning, natural computing, evolutionary architecture and software engineering. The research is centered around finding structure in data, recognizing anomalies during operation, adapting the learning architecture and (re-)organizing and evolving own learning processes at runtime against unforeseen situations. Furthermore, the whole approach will be conceptualized into a novel self-learning system to realize the integrated vision of unsupervised lifelong learning. ULEARN ultimately aims at autonomic computing systems contributing to tipping points in the next decade that will change our all lives. Different scenarios will be used to test and evaluate the research findings.

Funding scheme:

FRINATEK-Fri prosj.st. mat.,naturv.,tek