In this paper we consider a version of the zero-shot learning problem where seen class source and target domain data are provided. The goal during test-time is to accurately predict label an unseen instance based on revealed side information (e.g. attributes) for classes. Our method viewing each or as mixture proportions postulate that patterns have be similar if two instances belong same class. This perspective leads us source/target embedding functions map arbitrary into semantic space...

Zero-shot recognition (ZSR) deals with the problem of predicting class labels for target domain instances based on source side information (e.g. attributes) unseen classes. We formulate ZSR as a binary prediction problem. Our resulting classifier is class-independent. It takes an arbitrary pair and input predicts whether or not they come from same class, i.e. there match. model posterior probability match since it sufficient statistic propose latent probabilistic in this context. develop...

Anomalies in many video surveillance applications have local spatio-temporal signatures, namely, they occur over a small time window or spatial region. The distinguishing feature of these scenarios is that outside this anomalous region, activities appear normal. We develop probabilistic framework to account for such anomalies. show our admits elegant characterization optimal decision rules. A key insight the paper if anomalies are rules even when nominal behavior exhibits global and temporal...

The fundamental task of group testing is to recover a small distinguished subset items from large population while efficiently reducing the total number tests (measurements). key contribution this paper in adopting new information-theoretic perspective on problems. We formulate problem as channel coding/decoding and derive single-letter characterization for used identify defective set. Although focus primarily testing, our main result generally applicable other compressive sensing models....

In this paper we derive information theoretic performance bounds to sensing and reconstruction of sparse phenomena from noisy projections. We consider two settings: output noise models where the enters after projection input before projection. types distortion for reconstruction: support errors mean-squared errors. Our goal is relate number measurements, $m$, $\snr$, signal sparsity, $k$, level, $d$, dimension, $n$. in a worst-case setting. employ different variations Fano's inequality...

Non-line-of-sight (NLOS) propagation can severely degrade the reliability of communication and localisation accuracy in indoor ultra-wideband (UWB) 'location-aware' networks. Link adaptation NLOS bias mitigation techniques have respectively been proposed to alleviate these effects, but implicitly rely on ability accurately distinguish between LOS scenarios. A statistical identification technique based hypothesis-testing received signal parameters UWB channels is discussed. In contrast...

We consider the problem of classifying among a set M hypotheses via distributed noisy sensors. The sensors can collaborate over communication network and task is to arrive at consensus about event after exchanging messages. apply variant belief propagation as strategy for collaboration solution classification problem. show that message evolution be reformulated linear dynamical system, which primarily characterized by connectivity. centralized maximum posteriori (MAP) estimate almost always...

We consider the problem of detecting a small subset defective items from large set via non-adaptive "random pooling" group tests. both case when measurements are noiseless, and case2 noisy (the outcome each test may be independently faulty with probability q). Order-optimal results for these scenarios known in literature. give information-theoretic lower bounds on query complexity problems, provide corresponding computationally efficient algorithms that match up to constant factor. To best...

We explore a location based approach for behavior modeling and abnormality detection. In contrast to the conventional object where an may first be tagged, identified, classified, tracked, we proceed directly with event characterization at pixel(s) level on motion labels obtained from background subtraction. Since events are temporally spatially dependent, this calls techniques that account statistics of spatiotemporal events. Based labels, learn co-occurrence normal across space-time. For...

We explore a location based approach for behavior modeling and abnormality detection. In contrast to the conventional object where an may first be tagged, identified, classified, tracked, we proceed directly with event characterization at pixel(s) level on motion labels obtained from background subtraction. Since events are temporally spatially dependent, this calls techniques that account statistics of spatiotemporal events. Based labels, learn co-occurrence normal across space-time. For...

We consider some computationally efficient and provably correct algorithms with near-optimal sample complexity for the problem of noisy nonadaptive group testing. Group testing involves grouping arbitrary subsets items into pools. Each pool is then tested to identify defective items, which are usually assumed be sparse. randomly pooling measurements, where pools selected independently test outcomes. also a model measurements allow both false negative positive outcomes (and asymmetric noise,...

Background subtraction is a powerful mechanism for detecting change in sequence of images that finds many applications. The most successful background methods apply probabilistic models to intensities evolving time; nonparametric and mixture-of-Gaussians are but two examples. main difficulty designing robust algorithm the selection detection threshold. In this paper, we adapt threshold varying video statistics by means statistical models. addition model, introduce foreground model based on...

We propose a novel Generalized Zero-Shot learning (GZSL) method that is agnostic to both unseen images and semantic vectors during training. Prior works in this context map high-dimensional visual features the domain, which we believe contributes gap. To bridge gap, low-dimensional embedding of instances "visually semantic." Analogous data quantifies existence an attribute presented instance, components our prototypical part-type instance. In parallel, as thought experiment, quantify impact...

We present an approach to adaptively utilize deep neural networks in order reduce the evaluation time on new examples without loss of accuracy. Rather than attempting redesign or approximate existing networks, we propose two schemes that networks. first pose adaptive network scheme, where learn a system choose components be evaluated for each example. By allowing correctly classified using early layers exit, avoid computational associated with full network. extend this selection selects show...

In this paper, we propose training very deep neural networks (DNNs) for supervised learning of hash codes. Existing methods in context train relatively "shallow" limited by the issues arising back propagation (e.g. vanishing gradients) as well computational efficiency. We a novel and efficient algorithm inspired alternating direction method multipliers (ADMM) that overcomes some these limitations. Our decomposes process into independent layer-wise local updates through auxiliary variables....

Nonadaptive group testing involves grouping arbitrary subsets of n items into different pools. Each pool is then tested and defective are identified. A fundamental question minimizing the number pools required to identify at most d items. Motivated by applications in network tomography, sensor networks infection propagation, a variation problems on graphs formulated. Unlike conventional problems, each here must conform constraints imposed graph. For instance, can be associated with vertices...

As we move toward large-scale object detection, it is unrealistic to expect annotated training data, in the form of bounding box annotations around objects, for all classes at sufficient scale; therefore, methods capable unseen detection are required. We propose a novel zero-shot method based on an end-to-end model that fuses semantic attribute prediction with visual features boxes seen and classes. While utilize during training, our agnostic information test-time. Our retains efficiency...

In this paper we consider a version of the zero-shot learning problem where seen class source and target domain data are provided. The goal during test-time is to accurately predict label an unseen instance based on revealed side information (\eg attributes) for classes. Our method viewing each or as mixture proportions postulate that patterns have be similar if two instances belong same class. This perspective leads us source/target embedding functions map arbitrary into semantic space...