Open World
The world is almost unlimited and evolves constantly. Let us take machine learning into this world and go beyond evaluation on static train-test datasets.
Lifelong Learning
Humans never stop learning and profit from structured curricula. Let us equip machine learning systems with this ability and continuously grow their knowledge.
Wholistic Systems
Machine learning algorithms often get tailored to a specific focus. Let us investigate the complex interplays and exploit synergies between techniques.
About the Lab
The Open World Lifelong Learning (OWLL) Lab has been founded in March 2022 as a DEPTH independent research group within the Hessian Center for Artificial Intelligence (hessian.AI) and the state-funded research-cluster The Third Wave of Artificial Intelligence (3AI), both coordinated by Technical University of Darmstadt (TU Darmstadt).
OWLL is led by Martin Mundt, who is presently also a visiting professor at TU Darmstadt.
Our Vision
Whereas it is true that large progress in machine learning is being made, we are unfortunately also observing a natural effect of too strong extrapolation beyond the scientific findings. We posit that this may partially be a result of the way that machine learning has been, and is still, investigated today. We tend to pick a task, and gather some data for it, choose a statistical model to learn it, and later commonly conclude that our approach is successful, if it performs well on a set of data we have dedicated for testing. Often this process is repeated several times to improve the outcome by either tuning the model, a model-centric view, or by collecting or curating more data, a data-centric view. However, well-grounded real-world systems tend to require more than featuring the best performance number on some benchmark. Recent trends in machine learning have thus revived the idea for systems to be robust, systems to have human understandable failure modes, and systems to learn from data streams continually over time.
What is now missing to advance to the next stage is a realization that these elements are inevitably interconnected. Explaining the respective dependencies and leveraging emerging synergies is a key element to OWLL’s research on a path to open world lifelong learning. As the world is constantly changing, open world lifelong learning envisions to enable our next wave of AI systems to autonomously adapt to these changes. Our central hypothesis is that this can only be achieved if we incorporate a wholistic view, which incorporates a dynamic perspective on evolving data, adaptive models, and changing evaluation requirements, natively into the design process.
Our Objectives
OWLL aims to develop novel methods and interconnect subfields to create the next generation of AI systems that can learn independently in an open-ended world. Such systems can not only learn continuously, but also successfully recognize new situations and actively choose data to train on, while autonomously adapting in a robust and interpretable way.
Whereas we believe that this can only be achieved if we address open questions through a joint design process, we posit that the below individual properties are central elements.
Adaptive Models
Dynamic architectures that can automatically adapt to changing requirements of complexity over time
Knowledge Transfer
Optimization that enables transfer of existing knowledge to novel tasks, while avoiding forgetting of the already learned representations.
Active Data Choice
Active data selection processes to acknowledge distribution shifts and take into account the effects of ordering or task similarity.
Robust Predictions
Models that can confidently predict on what is known from training, but indicate high uncertainty for unseen and unknown concepts.