The mission of UvA-Bosch DELTA Lab (Deep Learning Technologies Amsterdam) is to perform world class research in the fundamentals of deep learning, with applications to automotive and computer vision. Researchers at UvA will collaborate with researchers at Bosch through an exchange program. Topics include: Reinforcement Learning, Bayesian Deep Learning, Data Efficient Deep Learning, Multi-Modal Deep Learning, Generative and Unsupervised Deep Learning, Model Compression, and Adversarial Deep Learning.

Project 1: Methods for Semi-supervised Learning and Active Labeling
Exploiting unlabeled data for a supervised learning problem and identifying the most informative subset of examples to be annotated by an expert.
Project 2: Methods for Robust Feature Learning
Learning robust features that remain maximally predictive even if the distribution of test data is very different from the distribution of training data.
Project 3: Calibrated Uncertainty Estimation
Providing reliable confidence intervals for deep neural network predictions.
Project 4: Methods for Multimodal Learning and Sensor Fusion
Combining multiple sources of information to improve prediction accuracy.
Project 5: Combining Generative Probabilistic Models with Deep Learning
Using probabilistic, possibly causal, graphical models, or complex simulators, to improve the accuracy of a classifier.
Project 6: Model Compression and Distillation
Maximally compressing the amount of bits necessary to store and execute a deep neural network while maintaining high accuracy.
Project 7: Reinforcement Learning and Planning
Using reinforcement learning to plan the actions of e.g. a car in traffic, given sensory information of its surroundings.
Project 8: Learning color-invariant bases
Learning robust, universally applicable color-invariants in the lower layers of CNN’s that facilitate image classification.
Project 9: Learning to follow objects over multiple cameras
Learning the characteristics of objects as observed from multiple camera’s images without a priori knowledge on the camera’s properties, their frames or the objects.
Project 10: Learning from images near the boundary of a class
Learning from adversarial examples or hard positive/negative examples and making classifiers perform robustly when confronted with adversarial examples.