TechTalks from event: Domain Adaptation Workshop: Theory and Application

We are still uploading slides and videos for this event. Please excuse any discrepancy.

Morning Session

Afternoon Session

  • Invited Speaker Authors: Dan Roth
  • History Dependent Domain Adaptation Authors: Allen Lavoie, Matthew Eric Otey, Nathan Ratliff and D. Sculley
    We study a novel variant of the domain adaptation problem, in which the loss function on test data changes due to dependencies on prior predictions. One important instance of this problem area occurs in settings where it is more costly to make a new error than to repeat a previous error. We propose several methods for learning effectively in this setting, and test them empirically on the real-world tasks of malicious URL classification and adversarial advertisement detection.
  • Domain Adaptation with Multiple Latent Domains Authors: Judy Hoffman, Kate Saenko, Brian Kulis and Trevor Darrell
    Domain adaptation is important for practical applications of supervised learning, as the distribution of inputs can differ significantly between available sources of training data and the test data in a particular target domain. Many domain adaptation methods have been proposed, yet very few of them deal with the case of more than one training domain; methods that do incorporate multiple domains assume that the separation into domains is known a priori, which is not always the case in practice. In this paper, we introduce a method for multi-domain adaptation with unknown domain labels, based on learning nonlinear crossdomain transforms, and apply it to image classification. Our key contribution is a novel version of constrained clustering; unlike many existing constrained clustering algorithms, ours can be shown to provably converge locally while satisfying all constraints. We present experiments on a commonly available image dataset.
  • Overfitting and Small Sample Statistics Authors: Dean Foster, Sham Kakade and Ruslan Salakhutdinov
    We study the prevalent problem when a test distribution differs from the training distribution. We consider a setting where our training set consists of a small number of sample domains, but where we have many samples in each domain. Our goal is to generalize to a new domain. For example, we may want to learn a similarity function using only certain classes of objects, but we desire that this similarity function be applicable to object classes not present in our training sample (e.g. we might seek to learn that “dogs are similar to dogs” even though images of dogs were absent from our training set). Our theoretical analysis shows that we can select many more features than domains while avoiding overfitting by utilizing data-dependent variance properties. We present a greedy feature selection algorithm based on using T-statistics. Our experiments validate this theory showing that our T-statistic based greedy feature selection is more robust at avoiding overfitting than the classical greedy procedure.
  • Cool world: domain adaptation of virtual and real worlds for human detection using active learning Authors: David V´azquez, Antonio M. L´opez, Daniel Ponsa and Javier Marin
    Image based human detection is of paramount interest for different applications. The most promising human detectors rely on discriminatively learnt classifiers, i.e., trained with labeled samples. However, labelling is a manual intensive task, especially in cases like human detection where it is necessary to provide at least bounding boxes framing the humans for training. To overcome such problem, in Marin et al. we have proposed the use of a virtual world where the labels of the different objects are obtained automatically. This means that the human models (classifiers) are learnt using the appearance of realistic computer graphics. Later, these models are used for human detection in images of the real world. The results of this technique are surprisingly good. However, these are not always as good as the classical approach of training and testing with data coming from the same camera and the same type of scenario. Accordingly, in Vazquez et al. we cast the problem as one of supervised domain adaptation. In doing so, we assume that a small amount of manually labeled samples from real-world images is required. To collect these labeled samples we use an active learning technique. Thus, ultimately our human model is learnt by the combination of virtual- and real-world labeled samples which, to the best of our knowledge, was not done before. Here, we term such combined space cool world. In this extended abstract we summarize our proposal, and include quantitative results from Vazquez et al. showing its validity.