Convolutional neural networks (CNNs) have shown great success in computer vision, approaching human-level performance when trained for specific tasks via application-specific loss functions. In this paper, we propose a method for augmenting and training CNNs so that their learned features are compositional. It encourages networks to form representations that disentangle objects from their surroundings and from each other, thereby promoting better generalization. Our method is agnostic to the specific details of the underlying CNN to which it is applied and can in principle be used with any CNN. As we show in our experiments, the learned representations lead to feature activations that are more localized and improve performance over non-compositional baselines in object recognition tasks.

Transfer learning allows knowledge to be extracted from auxiliary domains and be used to enhance learning in a target domain. For transfer learning to be successful, it is critical to find the similarity between auxiliary and target domains, even when such mappings are not obvious. In this paper, we present a novel algorithm for finding the structural similarity between two domains, to enable transfer learning at a structured knowledge level. In particular, we address the problem of how to learn a non-trivial structural similarity mapping between two different domains when they are completely different on the representation level. This problem is challenging because we cannot directly compare features across domains. Our algorithm extracts the structural features within each domain and then maps the features into the Reproducing Kernel Hilbert Space (RKHS), such that the "structural dependencies" of features across domains can be estimated by kernel matrices of the features within each domain. By treating the analogues from both domains as equivalent, we can transfer knowledge to achieve a better understanding of the domains and improved performance for learning. We validate our approach on synthetic and real-world datasets.

In pervasive computing, localizing a user in wireless indoor environments is an important yet challenging task. Among the state-of-art localization methods, fingerprinting is shown to be quite successful by statistically learning the signal to location relations. However, a major drawback for fingerprinting is that, it usually requires a lot of labeled data to train an accurate localization model. To establish a fingerprinting-based localization model in a building with many floors, we have to collect sufficient labeled data on each floor. This effort can be very burdensome. In this paper, we study how to reduce this calibration effort by only collecting the labeled data on one floor, while collecting unlabeled data on other floors. Our idea is inspired by the observation that, although the wireless signals can be quite different, the floor-plans in a building are similar. Therefore, if we co-embed these different floors' data in some common low-dimensional manifold, we are able to align the unlabeled data with the labeled data well so that we can then propagate the labels to the unlabeled data. We conduct empirical evaluations on real-world multi-floor data sets to validate our proposed method.

We consider feature extraction (dimensionality reduction) for compositional data, where the data vectors are constrained to be positive and constant-sum. In real-world problems, the data components (variables) usually have complicated "correlations" while their total number is huge. Such scenario demands feature extraction. That is, we shall de-correlate the components and reduce their dimensionality. Traditional techniques such as the Principle Component Analysis (PCA) are not suitable for these problems due to unique statistical properties and the need to satisfy the constraints in compositional data. This paper presents a novel approach to feature extraction for compositional data. Our method first identifies a family of dimensionality reduction projections that preserve all relevant constraints, and then finds the optimal projection that maximizes the estimated Dirichlet precision on projected data. It reduces the compositional data to a given lower dimensionality while the components in the lower-dimensional space are de-correlated as much as possible. We develop theoretical foundation of our approach, and validate its effectiveness on some synthetic and real-world datasets.

In multiple instance learning (MIL), how the instances determine the bag-labels is an essential issue, both algorithmically and intrinsically. In this paper, we show that the mechanism of how the instances determine the bag-labels is different for different application domains, and does not necessarily obey the traditional assumptions of MIL. We therefore propose an adaptive framework for MIL that adapts to different application domains by learning the domain-specific mechanisms merely from labeled bags. Our approach is especially attractive when we are encountered with novel application domains, for which the mechanisms may be different and unknown. Specifically, we exploit mixture models to represent the composition of each bag and an adaptable kernel function to represent the relationship between the bags. We validate on synthetic MIL datasets that the kernel function automatically adapts to different mechanisms of how the instances determine the bag-labels. We also compare our approach with state-of-the-art MIL techniques on real-world benchmark datasets.

Proportional data (normalized histograms) have been frequently occurring in various areas, and they could be mathematically abstracted as points residing in a geometric simplex. A proper distance metric on this simplex is of importance in many applications including classification and information retrieval. In this paper, we develop a novel framework to learn an optimal metric on the simplex. Major features of our approach include: 1) its flexibility to handle correlations among bins/dimensions; 2) widespread applicability without being limited to ad hoc backgrounds; and 3) a "real" global solution in contrast to existing traditional local approaches. The technical essence of our approach is to fit a parametric distribution to the observed empirical data in the simplex. The distribution is parameterized by affinities between simplex vertices, which is learned via maximizing likelihood of observed data. Then, these affinities induce a metric on the simplex, defined as the earth mover's distance equipped with ground distances derived from simplex vertex affinities.