University of California,
Domain Adaptation and Deep Learning for Large Scale Object Recognition and Detection
I'll review recent progress exploiting deep learned representations in domain adaptation settings. A central question in domain adaptation is whether domain shift is still a significant problem as the size of the source domain grows large. I'll present results that suggest that while deep representations are far more semantically general and thus inherently domain invariant, even with the largest available datasets one observes marked dataset bias that can be ameliorated with domain adaptation techniques. I'll also present our recently introduced adaptation approach to large scale object detection, which treats classification data as a source domain and detection data as a target domain. We adapt a novel 7.5K-category detection network based on the R-CNN framework, which is able to produce relatively accurate detections for each category in less than one second per frame on a single GPU. I'll conclude by discussing future directions in domain adaptation, focusing on new adaptation paradigms including continuous and hierarchically structured domains.
University of Southern California.
Kernel Methods for Domain Adaptation
The problem of domain adaptation occurs when the test data (of a target domain) and training data (of some source domain(s)) are generated by different distributions. It arises in a variety of computer vision applications.
In this talk, I will present some of our recent efforts on unsupervised domain adaptation using kernel methods. One cannot solve the domain adaptation problems given arbitrary source-target pairs. Instead, we have to explore the structures or properties in data, under which potentially successful solutions exist. Kernel methods are versatile in modeling such structures or properties. I will demonstrate several kernel methods ("kernel trick", discriminative multiple kernel learning, kernel embedding of distributions, etc.) which have been successfully used to model the structures of subspaces, landmarks, and latent domains.
Inst. of Science and Technology,
Learning with a Time-evolving Data Distribution
Domain adaptation studies the problem of machine learning when the data distribution at training time (source) differs from the data distribution at prediction time (target). In my talk I will introduce a special domain adaptation scenario, in which the mismatch is due to an underlying time-evolution of the data distribution. We have access to sample sets from earlier time steps, but what we are really interested in the behavior of the distribution in the future. To tackle this problem, I will discuss a method for learning an operator that can extrapolate the dynamics of such the data distribution. It relies on two recent machine learning techniques: the embedding of probability distributions into a reproducing kernel Hilbert space, and vector-valued regression.
Courant Institute of Mathematical Sciences, New York.
Recent Theoretical and Algorithmic Advances in Domain Adaptation
Domain adaptation is a challenging learning problem whose scenario does not match the assumptions commonly adopted in formal studies. A theory of adaptation has been developed over the past few years based on the notion of discrepancy, a key measure of the divergence between distributions, relevant to tasks such as domain adaptation and transfer learning. This talk will briefly outline several of the insights provided by that theory as well as an effective learning algorithm it suggests. It will further describe more recent theoretical and algorithmic advances that depart from the common fixed reweighting techniques and report the results of experiments demonstrating their benefits.
Joint work with Corinna Cortes (Google Research) and Andres Munoz Medina (Courant Institute of Mathematical Sciences).
Katholieke Universiteit Leuven.
Overcoming Dataset Bias: How Far are We from the Solution
First, I'll discuss a few extensions on top of the Subspace Alignment method for unsupervised domain adaptation, which we introduced at last ICCV. In particular, I'll describe a method to efficiently estimate the subspace dimensionality even when working in high-dimensional feature spaces, as well as a way to take the source class labels into account during the construction of the subspaces.
Next, I'll report on results of a study where we analyzed the interplay between image representations on the one hand (going from simple Bag-of-Words over Fisher Vectors to deep learning based features) and the effectiveness of domain adaptation on the other.
Based on these results, I'll move on to some more general discussion on domain adaptation, open challenges, and how far we are from solving the problem.