Research line leader: Cees Snoek (UvA)
R1 researches 3 scenarios for data-efficient deep learning: i) unsupervised, ii) semi-supervised, and iii) active learning. R1 challenges are:
- How to deep learn when labeled data is absent?
- How to deep learn when labeled data is limited?
- How to deep learn when labeled data is interactive?
Deep learning is successful when massive amounts of labeled data are available, as evidenced by astonishing results in speech recognition, machine translation and image categorization. For many problems, however, the precious labeled data may be scarce, hard to obtain or simply unavailable; surveillance, healthcare and finance are just three domains that suffer from data-deficiency. For general-purpose deep learning uptake, data-efficiency is a necessity.
The ultimate data-efficiency is to use no labeled data at all, known as unsupervised machine learning. This topic has been identified by many as a grand challenge in DL, unlikely to be resolved soon, as existing methods lack in scalability, theoretical justification, or both. Arguably, a more viable near-term solution is zero-shot learning, a methodology able to solve a task despite not having received any training examples of that task. Both paradigms are promising for data-efficient DL, warranting further research. An alternative solution to data-deficiency is to simulate, crowdsource or gamify labeled data. With a limited amount of supervised data, the semi-supervised learning paradigm is attractive as it learns to combine labeled and unlabeled data points by assuming data are smooth, clustered, or lie on a manifold. A special case is weakly-supervised learning where labels at a coarse granularity, say an image label, are exploited for fine-granularity tasks like spatial localization of objects. These techniques are used in traditional machine learning; its application in DL requires further research. If above methods fail, the last resort is to supplement the labeled data interactively using active learning. Its objective is to optimize a model within a limited time and annotation budget by selecting the most informative training instances. Rather than separating the representation from the learning, we study end-to-end active learning strategies.