Real-World Learning

Progress in artificial intelligence has been astonishing in the past decade. Cars self-driving on highways, machines beating go-masters, and cameras categorizing images in a pixel-precise fashion are now common place, thanks to data-and-label supervised deep learning. Despite the impressive advances, it is becoming increasingly clear that deep learning networks are heavily biased towards their training conditions and become brittle when deployed under real-world situations that differ from those perceived during learning in terms of data, labels and objectives. Simply scaling-up along all dimensions at training time seems a dead end, not only because of the compute, storage and ethical expenses, but especially as humans are easily able to generalize robustly in a data-efficient fashion. Several learning paradigms have been proposed to account for the limitations of deep learning with the i.i.d. assumption. Shifting data distributions are attacked by domain adaptation and domain generalization, changing label vocabularies are the topic of interest in zero-shot, open set and open world learning, while varying objectives are covered in meta-learning and continual learning regimes. However, there is as of yet no learning methodology that can dynamically learn to generalize and adapt across domains, labels and tasks simultaneously, and do so in a data-efficient fashion. This is the ambitious long-term goal of ‘real-world learning’. Prof. Cees Snoek (full professor in computer science at the University of Amsterdam) will present some initial approaches and results towards its objective.

Tuesday, 11-23-2021, 05:00 p.m. – 06:00 p.m., virtual event

Further details:

This is an event of I-AIDA – International Artificial Intelligence Doctoral Academy.