Invited speakers
University of Southern California, USA
Title: What kind of learning structures do we need for zero/few-shot learning?
Abstract: Zero/few-shot (transfer) learning and domain adaptation are important settings for learning from small or limited visual data. In this talk, I will summarize a few recent work we have done in these directions. In particular, we are interested in the following question: what are the fundamental modeling assumptions (and resulting learning architectures) we need for successful learning? While we do not have a complete answer to that yet, I will give a few successful and surprising examples from our and others' work.
Talk slides: TBA
The University of Tokyo, Japan
Title: Domain Adaptation for Image Generation and Object Detection
Abstract: Deep neural networks have boosted recognition accuracy, but still rely on a large amount of high quality labeled data. The domain adaptation (DA), which transfers knowledge from label-rich domains to a label-scarce domain, is one of the promising techniques to solve this challenging problem. So far, a lot of studies in DA focus on object recognition tasks, but there aren't so many studies about object detection and image generation. In this talk, we introduce novel domain adaptation methods for object detection and image generation. Specifically, we propose a technique for detector adaptation based on strong local alignment and weak global alignment between the source and target domains. Then, we present a novel knowledge transfer method for GANs that can train generative models from extremely small-sized data by only adapting the batch statistics parameters. Besides, we carefully examine the theoretical aspect of unsupervised DA (UDA) and consider how to improve UDA accuracy between domains with a large discrepancy.
Mapillary Research, Austria
Title: Domain Alignment Layers for Unsupervised Domain Adaptation
Abstract: Classifiers trained on given databases perform poorly when tested on data acquired in different settings. This is explained in domain adaptation through a shift among distributions of the source and target domains. Attempts to align them have traditionally resulted in works reducing the domain shift by introducing appropriate loss terms, measuring the discrepancies between source and target distributions, in the objective function.
In this talk I will introduce Domain Alignment Layers (DALs), which provide an alternative, simple and effective way of addressing domain adaptation problems within modern neural networks. As opposed to previous works, DALs deal with the domain shift by aligning explicitly in the network the distribution of domain-specific features to a domain-agnostic target distribution. Moreover, our method can be adapted to automatically learn the degree of feature alignment required at different levels of the deep network, sidestepping the problem of deciding which layers should undergo domain adaptation.
Thorough experiments on different public benchmarks, in the unsupervised setting, confirm the power of our approach. Finally, I will present some additional results on domain generalization, predictive and continuous domain adaptation, which make use of DALs.
Naver Labs Europe, France
Title: New Trends in Visual Domain Adaptation
Abstract: The aim of this talk will be to give an overview of visual domain adaptation methods. First, I will briefly recall historical shallow methods and show how they can be used to improve long-term for visual localization. Then, I will discuss different manners to exploit deep convolutional architectures in visual domain adaptation, focusing on image categorization. In the last part of my talk, I will present deep domain adaptation models proposed for tasks beyond image classification, and in particular for semantic segmentation.