Search

Resampling is an important signature of manipulated images. In this paper, wepropose two methods to detect and localize image manipulations based on acombination of resampling features and deep learning. In the first method, theRadon transform of resampling features are computed on overlapping imagepatches. Deep learning classifiers and a Gaussian conditional random fieldmodel are then used to create a heatmap. Tampered regions are located using aRandom Walker segmentation method. In the second method, resampling featurescomputed on overlapping image patches are passed through a Long short-termmemory (LSTM) based network for classification and localization. We compare theperformance of detection/localization of both these methods. Our experimentalresults show that both techniques are effective in detecting and localizingdigital image forgeries.

When modeling geo-spatial data, it is critical to capture spatialcorrelations for achieving high accuracy. Spatial Auto-Regression (SAR) is acommon tool used to model such data, where the spatial contiguity matrix (W)encodes the spatial correlations. However, the efficacy of SAR is limited bytwo factors. First, it depends on the choice of contiguity matrix, which istypically not learnt from data, but instead, is assumed to be known apriori.Second, it assumes that the observations can be explained by linear models. Inthis paper, we propose a Convolutional Neural Network (CNN) framework to modelgeo-spatial data (specifi- cally housing prices), to learn the spatialcorrelations automatically. We show that neighborhood information embedded insatellite imagery can be leveraged to achieve the desired spatial smoothing. Anadditional upside of our framework is the relaxation of linear assumption onthe data. Specific challenges we tackle while implementing our frameworkinclude, (i) how much of the neighborhood is relevant while estimating housingprices? (ii) what is the right approach to capture multiple resolutions ofsatellite imagery? and (iii) what other data-sources can help improve theestimation of spatial correlations? We demonstrate a marked improvement of 57%on top of the SAR baseline through the use of features from deep neuralnetworks for the cities of London, Birmingham and Liverpool.

Large collections of videos are grouped into clusters by a topic keyword,such as Eiffel Tower or Surfing, with many important visual concepts repeatingacross them. Such a topically close set of videos have mutual influence on eachother, which could be used to summarize one of them by exploiting informationfrom others in the set. We build on this intuition to develop a novel approachto extract a summary that simultaneously captures both importantparticularities arising in the given video, as well as, generalities identifiedfrom the set of videos. The topic-related videos provide visual context toidentify the important parts of the video being summarized. We achieve this bydeveloping a collaborative sparse optimization method which can be efficientlysolved by a half-quadratic minimization algorithm. Our work builds upon theidea of collaborative techniques from information retrieval and naturallanguage processing, which typically use the attributes of other similarobjects to predict the attribute of a given object. Experiments on twochallenging and diverse datasets well demonstrate the efficacy of our approachover state-of-the-art methods.

Person re-identification is an open and challenging problem in computervision. Existing approaches have concentrated on either designing the bestfeature representation or learning optimal matching metrics in a static settingwhere the number of cameras are fixed in a network. Most approaches haveneglected the dynamic and open world nature of the re-identification problem,where a new camera may be temporarily inserted into an existing system to getadditional information. To address such a novel and very practical problem, wepropose an unsupervised adaptation scheme for re-identification models in adynamic camera network. First, we formulate a domain perceptivere-identification method based on geodesic flow kernel that can effectivelyfind the best source camera (already installed) to adapt with a newlyintroduced target camera, without requiring a very expensive training phase.Second, we introduce a transitive inference algorithm for re-identificationthat can exploit the information from best source camera to improve theaccuracy across other camera pairs in a network of multiple cameras. Extensiveexperiments on four benchmark datasets demonstrate that the proposed approachsignificantly outperforms the state-of-the-art unsupervised learning basedalternatives whilst being extremely efficient to compute.

The mass adoption of plug-in electric vehicles (PEVs) requires the deploymentof public charging stations. Such facilities are expected to employ distributedgeneration and storage units to reduce the stress on the grid and boostsustainable transportation. While prior work has made considerable progress inderiving insights for understanding the adverse impacts of PEV chargings andhow to alleviate them, a critical issue that affects the accuracy is the lackof real world PEV data. As the dynamics and pertinent design of such chargingstations heavily depend on actual customer demand profile, in this paper wepresent and evaluate the data obtained from a $17$ node charging networkequipped with Level $2$ chargers at a major North American University campus.The data is recorded for $166$ weeks starting from late $2011$. The resultindicates that the majority of the customers use charging lots to extend theirdriving ranges. Also, the demand profile shows that there is a tremendousopportunity to employ solar generation to fuel the vehicles as there is acorrelation between the peak customer demand and solar irradiation. Also, weprovided a more detailed data analysis and show how to use this information indesigning future sustainable charging facilities.

The capacity of next generation of cellular networks using femtocaches isstudied when multihop communications and decentralized cache placement areconsidered. We show that the storage capability of future network UserTerminals (UT) can be effectively used to increase the capacity in randomdecentralized uncoded caching. We further propose a random decentralized codedcaching scheme which achieves higher capacity results than the randomdecentralized uncoded caching. The result shows that coded caching which issuitable for systems with limited storage capabilities can improve the capacityof cellular networks by a factor of log(n) where n is the number of nodesserved by the femtocache.

Tensor contractions constitute a key computational ingredient of numericalmulti-linear algebra. However, as the order and dimension of tensors grow, thetime and space complexities of tensor-based computations grow quickly. Existingapproaches for tensor contractions typically involves explicit copy andtranspose operations. In this paper, we propose and evaluate a new BLAS-likeprimitive STRIDEDBATCHEDGEMM that is capable of performing a wide range oftensor contractions on CPU and GPU efficiently. Through systematicbenchmarking, we demonstrate the advantages of our approach over conventionalapproaches. Concretely, we implement the Tucker decomposition and show thatusing our kernels yields 100x speedup as compared to the implementation usingexisting state-of-the-art libraries.

A non-orthogonal multiple access (NOMA) approach to user signal powerallocation called Fair-NOMA is introduced. Fair-NOMA is the application of NOMAin such a way that two mobile users have the opportunity to always achieve atleast the information capacity they can achieve by using orthogonal multipleaccess (OMA), regardless of the user selection criteria, making it suitable forimplementation using any current or future scheduling paradigms. Given thiscondition, the bounds of the power allocation coefficients are derived asfunctions of the channel gains of the two mobile users. The NOMA powerallocation is analyzed for two scheduled users that are selected randomly withi.i.d. channel gains. The capacity improvements made by each user and the sumcapacity improvement are derived.

We present an approach to constructing a neuromorphic device that responds tolanguage input by producing neuron spikes in proportion to the strength of theappropriate positive or negative emotional response. Specifically, we perform afine-grained sentiment analysis task with implementations on two differentsystems: one using conventional spiking neural network (SNN) simulators and theother one using IBM's Neurosynaptic System TrueNorth. Input words are projectedinto a high-dimensional semantic space and processed through a fully-connectedneural network (FCNN) containing rectified linear units trained viabackpropagation. After training, this FCNN is converted to a SNN bysubstituting the ReLUs with integrate-and-fire neurons. We show that there ispractically no performance loss due to conversion to a spiking network on asentiment analysis test set, i.e. correlations between predictions and humanannotations differ by less than 0.02 comparing the original DNN and its spikingequivalent. Additionally, we show that the SNN generated with this techniquecan be mapped to existing neuromorphic hardware -- in our case, the TrueNorthchip. Mapping to the chip involves 4-bit synaptic weight discretization andadjustment of the neuron thresholds. The resulting end-to-end system can take auser input, i.e. a word in a vocabulary of over 300,000 words, and estimate itssentiment on TrueNorth with a power consumption of approximately 50 uW.

We apply social ℓ-norms for the first time to the problem of hyperspectral unmixing while modeling spectral variability. These norms are built with inter-group penalties which are combined in a global intra-group penalization that can enforce selection of entire endmember bundles; this results in the selection of a few representative materials even in the presence of large endmembers bundles capturing each material's variability. We demonstrate improvements quantitatively on synthetic data and qualitatively on real data for three cases of social norms: group, elitist, and a fractional social norm, respectively. We find that the greatest improvements arise from using either the group or fractional flavor.

In recent years the field of neuromorphic low-power systems that consumeorders of magnitude less power gained significant momentum. However, theirwider use is still hindered by the lack of algorithms that can harness thestrengths of such architectures. While neuromorphic adaptations ofrepresentation learning algorithms are now emerging, efficient processing oftemporal sequences or variable length-inputs remain difficult. Recurrent neuralnetworks (RNN) are widely used in machine learning to solve a variety ofsequence learning tasks. In this work we present a train-and-constrainmethodology that enables the mapping of machine learned (Elman) RNNs on asubstrate of spiking neurons, while being compatible with the capabilities ofcurrent and near-future neuromorphic systems. This "train-and-constrain" methodconsists of first training RNNs using backpropagation through time, thendiscretizing the weights and finally converting them to spiking RNNs bymatching the responses of artificial neurons with those of the spiking neurons.We demonstrate our approach by mapping a natural language processing task(question classification), where we demonstrate the entire mapping process ofthe recurrent layer of the network on IBM's Neurosynaptic System "TrueNorth", aspike-based digital neuromorphic hardware architecture. TrueNorth imposesspecific constraints on connectivity, neural and synaptic parameters. Tosatisfy these constraints, it was necessary to discretize the synaptic weightsand neural activities to 16 levels, and to limit fan-in to 64 inputs. We findthat short synaptic delays are sufficient to implement the dynamical (temporal)aspect of the RNN in the question classification task. The hardware-constrainedmodel achieved 74% accuracy in question classification while using less than0.025% of the cores on one TrueNorth chip, resulting in an estimated powerconsumption of ~17 uW.

The problem of finding code distance has been long studied for the genericensembles of linear codes and led to several algorithms that substantiallyreduce exponential complexity of this task. However, no asymptotic complexitybounds are known for distance verification in other ensembles of linear codes.Our goal is to re-design the existing generic algorithms of distanceverification and derive their complexity for LDPC codes. We obtain newcomplexity bounds with provable performance expressed in terms of theerasure-correcting thresholds of long LDPC codes. These bounds exponentiallyreduce complexity estimates known for linear codes.

Spike-timing-dependent plasticity (STDP) incurs both causal and acausalsynaptic weight updates, for negative and positive time differences betweenpre-synaptic and post-synaptic spike events. For realizing such updates inneuromorphic hardware, current implementations either require forward andreverse lookup access to the synaptic connectivity table, or rely onmemory-intensive architectures such as crossbar arrays. We present a novelmethod for realizing both causal and acausal weight updates using only forwardlookup access of the synaptic connectivity table, permitting memory-efficientimplementation. A simplified implementation in FPGA, using a single timervariable for each neuron, closely approximates exact STDP cumulative weightupdates for neuron refractory periods greater than 10 ms, and reduces to exactSTDP for refractory periods greater than the STDP time window. Compared toconventional crossbar implementation, the forward table-based implementationleads to substantial memory savings for sparsely connected networks supportingscalable neuromorphic systems with fully reconfigurable synaptic connectivityand plasticity.

This paper proposes models of learning process in teams of individuals whocollectively execute a sequence of tasks and whose actions are determined byindividual skill levels and networks of interpersonal appraisals and influence.The closely-related proposed models have increasing complexity, starting with acentralized manager-based assignment and learning model, and finishing with asocial model of interpersonal appraisal, assignments, learning, and influences.We show how rational optimal behavior arises along the task sequence for eachmodel, and discuss conditions of suboptimality. Our models are grounded inreplicator dynamics from evolutionary games, influence networks frommathematical sociology, and transactive memory systems from organizationscience.

In this work, an investigation of the effect of mechanical preloads on the hysteresis loop of composite soft magnetorheological elastomer (MRE) was carried out. MRE is a “smart” composite material that consists of magnetically permeable particles in a non-magnetic polymeric elastomer. When subjected to an external magnetic field, a large deformational change occurs in the mechanical properties of these materials. Due to their coupled magnetomechanical response, these materials have been found suitable for various engineering applications. Inspired by experimental work, we present a model of the effect of mechanical preloads on the magnetization response of MRE based on a general continuum formulation. Using the Jiles - Atherton (JA) model parameters derived from the fitting of experimental measurement, the hysteresis loop of isotropic MRE was numerically resolved, which was then coupled to mechanical fields based on an energetically constitutive model valid for finitely strained MREs. Simulation analysis is performed for uniaxial stresses parallel to the direction of the applied magnetic field. For the applied tensile and compressive stresses, only a small change is observed in the hysteresis loop of these materials. Additionally, microscale modeling of the magnetization behavior of the isotropic MRE based on experimental results was performed. Considering the interaction between the magnetic particles, the magnetic and mechanical fields are resolved explicitly inside the composite material. A computational homogenization scheme was utilized to relate the microscopic behavior to the effective macroscopic properties of the MRE. In principle, the predicted effective magnetization behavior is observed to agree with the measured hysteresis loop of MRE materials.

In this paper we present a sufficient condition that guaranteesidentifiability of linear network dynamic systems exhibiting continuous-timeweighted consensus protocols with acyclic structure. Each edge of theunderlying network graph $\mathcal G$ of the system is defined by a constantparameter, referred to as the weight of the edge, while each node is defined bya scalar state whose dynamics evolve as the weighted linear combination of itsdifference with the states of its neighboring nodes. Following the classicaldefinitions of identifiability and indistinguishability, we first derive acondition that ensure the identifiability of the edge weights of $\mathcal G$in terms of the associated transfer function. Using this characterization, wepropose a sensor placement algorithm that guarantees identifiability of theedge weights. We describe our results using several illustrative examples.

Abstract Background It’s a very urgent task to identify cancer genes that enables us to understand the mechanisms of biochemical processes at a biomolecular level and facilitates the development of bioinformatics. Although a large number of methods have been proposed to identify cancer genes at recent times, the biological data utilized by most of these methods is still quite less, which reflects an insufficient consideration of the relationship between genes and diseases from a variety of factors. Results In this paper, we propose a two-rounds random walk algorithm to identify cancer genes based on multiple biological data (TRWR-MB), including protein-protein interaction (PPI) network, pathway network, microRNA similarity network, lncRNA similarity network, cancer similarity network and protein complexes. In the first-round random walk, all cancer nodes, cancer-related genes, cancer-related microRNAs and cancer-related lncRNAs, being associated with all the cancer, are used as seed nodes, and then a random walker walks on a quadruple layer heterogeneous network constructed by multiple biological data. The first-round random walk aims to select the top score k of potential cancer genes. Then in the second-round random walk, genes, microRNAs and lncRNAs, being associated with a certain special cancer in corresponding cancer class, are regarded as seed nodes, and then the walker walks on a new quadruple layer heterogeneous network constructed by lncRNAs, microRNAs, cancer and selected potential cancer genes. After the above walks finish, we combine the results of two-rounds RWR as ranking score for experimental analysis. As a result, a higher value of area under the receiver operating characteristic curve (AUC) is obtained. Besides, cases studies for identifying new cancer genes are performed in corresponding section. Conclusion In summary, TRWR-MB integrates multiple biological data to identify cancer genes by analyzing the relationship between genes and cancer from a variety of biological molecular perspective.

We consider the problem of recovering block-sparse signals whose structuresare unknown \emph{a priori}. Block-sparse signals with nonzero coefficientsoccurring in clusters arise naturally in many practical scenarios. However, theknowledge of the block structure is usually unavailable in practice. In thispaper, we develop a new sparse Bayesian learning method for recovery ofblock-sparse signals with unknown cluster patterns. Specifically, apattern-coupled hierarchical Gaussian prior model is introduced to characterizethe statistical dependencies among coefficients, in which a set ofhyperparameters are employed to control the sparsity of signal coefficients.Unlike the conventional sparse Bayesian learning framework in which eachindividual hyperparameter is associated independently with each coefficient, inthis paper, the prior for each coefficient not only involves its ownhyperparameter, but also the hyperparameters of its immediate neighbors. Indoing this way, the sparsity patterns of neighboring coefficients are relatedto each other and the hierarchical model has the potential to encouragestructured-sparse solutions. The hyperparameters, along with the sparse signal,are learned by maximizing their posterior probability via anexpectation-maximization (EM) algorithm. Numerical results show that theproposed algorithm presents uniform superiority over other existing methods ina series of experiments.

This paper addresses the problem of collaborative tracking of dynamic targetsin wireless sensor networks. A novel distributed linear estimator, which is aversion of a distributed Kalman filter, is derived. We prove that the filter ismean square consistent in the case of static target estimation. When largesensor networks are deployed, it is common that the sensors do not have goodknowledge of their locations, which affects the target estimation procedure.Unlike most existing approaches for target tracking, we investigate theperformance of our filter when the sensor poses need to be estimated by anauxiliary localization procedure. The sensors are localized via a distributedJacobi algorithm from noisy relative measurements. We prove strong convergenceguarantees for the localization method and in turn for the joint localizationand target estimation approach. The performance of our algorithms isdemonstrated in simulation on environmental monitoring and target trackingtasks.

The design and performance of a fully-synchronous multi-GHz analog transientwaveform recorder I.C. ("SST") with fast and flexible trigger capabilities ispresented. The SST's objective is to provide multi-GHz sample rates withintrinsically-stable timing, Nyquist-rate sampling and high trigger bandwidth,wide dynamic range and simple operation. Containing 4 channels of 256 samplesper channel, the SST is fabricated in an inexpensive 0.25 micrometer CMOSprocess and uses a high-performance package that is 8 mm on a side. It has a1.9V input range on a 2.5V supply, exceeds 12 bits of dynamic range, and uses~128 mW while operating at 2 G-samples/s and full trigger rates. With astandard 50 Ohm input source, the SST exceeds ~1.5 GHz -3 dB bandwidth. TheSST's internal sample clocks are generated synchronously via a shift registerdriven by an external LVDS oscillator running at half the sample rate (e.g., a1 GHz oscillator yields 2 G-samples/s). Because of its purely-digitalsynchronous nature, the SST has ps-level timing uniformity that is independentof sample frequencies spanning over 6 orders of magnitude: from under 2 kHz toover 2 GHz. Only three active control lines are necessary for operation: Reset,Start/Stop and Read-Clock. When operating as common-stop device, the time ofthe stop, modulo 256 relative to the start, is read out along with the sampledsignal values. Each of the four channels integrates dual-threshold triggercircuitry with windowed coincidence features. Channels can discriminate signalswith ~1mV RMS resolution at >600 MHz bandwidth.