On 6th of October between 12:00 and 13:00, ‘ ICAI: The Labs’ will have a meetup on AI for computer vision in the Netherlands. Two labs each present their work and discuss challenges and developments in this field.
QUVA Lab is the collaboration between Qualcomm and the University of Amsterdam. The mission of the QUVA-lab is to perform world-class research on deep vision. Such vision strives to automatically interpret with the aid of deep learning what happens where, when and why in images and video.
12.00 (noon): Opening
12:05 Introduction of the QUVA Lab by Yuki Asano
12:10 QUVA Lab: Philip Lippe about: “CITRUS: Causal Identifiability from Temporal Sequences with Interventions”
12:25 Introduction Lab 2 TBC
12:30 Presentation Lab 2 TBC
12.45 Discussion of what’s next in AI in Computer Vision
Understanding the underlying causal factors of a dynamical system is a crucial step towards agents reasoning in complex environments. Recent progress in causal representation learning and non-linear ICA showed that, in certain situations like factors being independent, it is possible to uncover such factors from observations. We seek to extend this line of work to a temporal domain where interventions have been performed on a subset of variables in between time steps, which resembles the setting of an agent taking actions over time. We propose CITRUS, a variational autoencoder framework which, with the help of interventional data, can identify the causal variables up to their intervention-dependent aspects and all remaining information. Further, CITRUS can be extended to pretrained autoencoders, opening up future research areas of simulation-to-real-world generalization for causal learning. Evaluated on visually complex image sequences with non-linear relations among causal variables, the method is capable to disentangle and recover the underlying causal variables.