Metric learning requires the identification of far-apart similar pairs and close dissimilar pairs during training, and this is difficult to achieve with unlabeled data because pairs are typically assumed to be similar if they are close. We present a novel metric learning method which circumvents this issue by identifying hard negative pairs as those which obtain dissimilar labels via label propagation (LP), when the edge linking the pair of data is removed in the affinity matrix. In so doing, the negative pairs can be identified despite their proximity, and we are able to utilize this information to significantly improve LP’ s ability to identify far-apart positive pairs and close negative pairs. This results in a considerable improvement in semi-supervised metric learning performance as evidenced by recall, precision and Normalized Mutual Information (NMI) performance metrics on Content-based Information Retrieval (CBIR) applications.

We present a supervised semantic hashing framework, named Label-Consistent Generalized Hashing (LCGH). The main novelty of LCGH is the explicit retention of information which may be irrelevant in training, but possibly useful for generalizing to unseen test classes. This is in stark contrast to typical semantic hashing methods which seek to remove redundant feature information from their hash codes in order to maximize the margin between hash codes of dissimilar data. This typical strategy leaves hash codes narrowly viable for discerning between training classes, and inadequate in discriminating between unseen test classes. Instead of limiting the information content of hash codes to those provided by the training labels, LCGH enhances its codes with information content from both supervised and unsupervised sources, improving their ability to discriminate across a wider range of data. To do so, LCGH builds upon the foundation of first agreeing with the provided training labels (label-consistency) and then incorporating possibly useful information using a reconstruction loss. In this way, LCGH respects the reliably given label information before exploring the addition of possibly useful ones. The outcome is a hashing scheme with slightly weaker within-domain (training and test classes are the same) retrieval performance, but much stronger cross-domain (training and test classes are disjoint) performance.

Three of the key challenges in semi-supervised metric learning are the difficulty in sampling loss-producing triplets, the difficulty in locating similar data which are faraway from the anchor points, and the difficulty in making the model robust to noisy predicted pseudolabels. We propose a method which allows the use of class-representative anchors (proxies), and avoids the computational costs associated with triplet sampling. Our new semi-supervised metric learning method propagates labels along mutual nearest neighbor pairs, so that faraway similar data can be drawn to the anchors, while data which are not along these paths (and hence not on the same manifold as the anchors) can be pushed away from these anchors. By assessing the number of different labels which were propagated to the same point, we obtain an estimate of the probability that our prediction of the pseudolabel is accurate, and hence able to attenuate the effect of uncertain pseudolabels on our model by factoring in the confidence of these predictions. We show the superiority of our method over various state-of-the-art methods on four diverse public datasets.

We present a supervised semantic hashing method which uses a variational autoencoder to represent each database image sample as a product Bernoulli distribution. We show that the probability parameters approach extreme values during training, allowing them to be used directly as hash bits. We show how our method allows balanced bits to be directly specified, and is superior to state-of-the-art methods across four datasets.

Cost-Sensitive Online Classification is recently proposed to directly online optimize two well-known cost-sensitive measures: (i) maximization of weighted sum of sensitivity and specificity, and (ii) minimization of weighted misclassification cost. However, the previous existing learning algorithms only utilized the first order information of the data stream. This is insufficient, as recent studies have proved that incorporating second order information could yield significant improvements on the prediction model. Hence, we propose a novel cost-sensitive online classification algorithm with adaptive regularization. We theoretically analyzed the proposed algorithm and empirically validated its effectiveness with extensive experiments. We also demonstrate the application of the proposed technique for solving several online anomaly detection tasks, showing that the proposed technique could be an effective tool to tackle cost-sensitive online classification tasks in various application domains.