The human brain’s computer‐assisted prognosis (CAP) system relies heavily on the self‐regulating characterization of tumors. Despite being extensively researched, the classification of brain tumors into meningioma, glioma, and pituitary types using magnetic resonance (MR) images presents significant challenges. Although biopsies are currently the gold standard for evaluating tumors, the need for noninvasive and accurate methods to grade brain tumors is increasing due to the risks associated with invasive biopsies. The objective is to introduce a noninvasive brain tumor grading system based on MR imaging (MRI) and deep learning (DL) utilizing probabilistic selection techniques. In the proposed method, the best three of the seven state‐of‐the‐art deep convolutional networks are chosen after extensive experimentation and combined with a probabilistic selection technique to enhance the overall performance of the proposed classification model. The results elucidate that the proposed model successfully classifies the tumor types into Glioma, Meningioma, and Pituitary achieving a sensitivity of 0.928, 0.939, and 0.992, respectively for each tumor type. Also, the precision in classifying the tumor classes is attained as 0.969, 0.932, and 0.957, respectively claiming an accuracy of 0.966, 0.956, and 0.983 for each of the three classes. The proposed model achieved an overall classification accuracy of 96.06%, surpassing the state‐of‐the‐art advanced and sophisticated techniques. Extensive experiments were performed on brain MRI datasets to demonstrate the enhanced performance of the proposed approach. The suggested probabilistic selection technique yielded promising classification results for brain tumors and exhibited the potential to leverage the strengths of various models.

One of the most common approaches through which people communicate is facial expressions. A large number of features documented in the literature were created by hand, with the goal of overcoming specific challenges such as occlusions, scale, and illumination variations. These classic methods are then applied to a dataset of facial images or frames in order to train a classifier. The majority of these studies perform admirably on datasets of images shot in a controlled environment, but they struggle with more difficult datasets (FER-2013) that have higher image variation and partial faces. The nonuniform features of the human face as well as changes in lighting, shadows, facial posture, and direction are the key obstacles. Techniques of deep learning have been studied as a set of methodologies for gaining scalability and robustness on new forms of data. In this paper, we look at how well-known deep learning techniques (e.g. GoogLeNet, AlexNet) perform when it comes to facial expression identification, and propose an enhanced hybrid deep learning model based on STN for facial emotion recognition, which gives the best feature extraction and classification in one go and maximizes the accuracy for a large number of samples on FERG, JAFFE, FER-2013, and CK+ datasets. It is capable of focusing on the main parts of the face and attaining extensive development over preceding fashions on the FERG, JAFFE, CK+ datasets, and the more challenging one namely FER-2013.

Anomalies in network traffic are often detected using machine learning techniques, such a Artificial Neural Networks, Self-Organizing Maps, k-Nearest Neighbors, or Principal Component Analysis. These techniques are built upon certain predetermined features that are believed to be useful in detecting anomalies. Many researchers are using graph-based features, such as betweenness centrality or eigenvector centrality. The choice of these particular features is due to the assumption that they can be used to accurately predict an anomaly in the flow of traffic. However, there appears to be no solid foundation for these assumptions. This work investigates edge centralities and how accurately they predict anomalies using netflow data. We propose to use known traits of different network interactions to identify how information will flow. We will then predict which measures of centrality should be most applicable to these particular flows. Finally, using public cybersecurity data sets, we will investigate which measures of edge centrality accurately identify anomalies as outliers then make comparisons with our predictions. Ideally, this will allow us to choose graph-based features that are highly efficient in anomaly detection.

The Department of Defense currently maintains a network known as the Defense Research Engineering Network (DREN), which provides various Department of Defense (DoD) sites across the nation connectivity to HPC resource centers. To ensure the security of the DREN system, a defense system known as the Cybersecurity Environment for Detection, Analysis, and Reporting (CEDAR) was created. CEDAR contains a variety of cybersecurity sensors, which constantly monitor and record real time network activity on the DREN. Over time, CEDAR has accumulated massive quantities of valuable cybersecurity data, which necessitates a form of automation in the process of reviewing this data. We propose the application of deep learning techniques to CEDAR data in an attempt to automatically detect potentially malicious activity in a more agile and adaptable manner. These deep learning techniques can be carried out in a high performance computing (HPC) environment, allowing for the rapid utilization of large amounts of data. Our most effective model is able to classify CEDAR alerts as malicious with an accuracy sufficient to greatly reduce human analyst workloads.

Road detection is a difficult task because roads are complex objects lacking a cohesive shape; as such, detection is typically limited to satellite images of urban environments. Detecting complex roads beyond asphalt, such as dirt or gravel, in complex environments, such as a desert, poses a more significant challenge. This report describes the implementation of a convolutional neural network to segmentize roads across a variety of types in a variety of environments from aerial images by directly comparing the image to a corresponding ground-truth. Current results are favorable for complex roads but the detector needs more work for environments where roads may take up most of the image.

Object detection, though gaining popularity, has largely been limited to detection from the ground or from satellite imagery. Aerial images, where the target may be obfuscated from the environmental conditions, angle-of-attack, and zoom level, pose a more significant challenge to correctly detect targets in. This paper describes the implementation of a regional convolutional neural network to locate and classify objects across several categories in complex, aerial images. Our current results show promise in detecting and classifying objects. Further adjustments to the network and data input should increase the localization and classification accuracies.

Humans can easily perform tasks that use vision and language jointly, such as describing a scene and answering questions about objects in the scene and how they are related. Image captioning and visual question & answer are two popular research tasks that have emerged from advances in deep learning and the availability of datasets that specifically address these problems. However recent work has shown that deep learning based solutions to these tasks are just as brittle as solutions for only vision or only natural language tasks. Image captioning is vulnerable to adversarial perturbations; novel objects, which are not described in training data, and contextual biases in training data can degrade performance in surprising ways. For these reasons, it is important to find ways in which general-purpose knowledge can guide connectionist models. We investigate challenges to integrate existing ontologies and knowledge bases with deep learning solutions, and possible approaches for overcoming such challenges. We focus on geo-referenced data such as geo-tagged images and videos that capture outdoor scenery. Geo-knowledge bases are domain specific knowledge bases that contain concepts and relations that describe geographic objects. This work proposes to increase the robustness of automatic scene description and inference by leveraging geo-knowledge bases along with the strengths of deep learning for visual object detection and classification.

Indexing research articles in scientific publications can be arduous. The authors tag their articles by the topics or domains relevant to their research. A publication's organizers may tag them by the broad topics of the specific publication. A third-party may index or tag these articles based on their subject knowledge. Hence indexing of articles can be uneven due to inconsistencies in area knowledge by third-parties or the niche topic representation by the authors. Publications may have schemes in place for indexing or tagging the articles but such schemes cannot keep up with the continuously changing landscape of research. These schemes may need to be updated with newer topics or domains being churned out by the state of the art research. Our technique endeavors to address this problem. We present a methodology to find similarity among domains extracted from the content of research papers, and cluster related domains. Analysis of these clusters provides insights into how the existing indexing schemes may be enhanced by adding newer domains.

Cloud Radio Access Networks (Cloud RAN) is an emerging architectural paradigm that attempts to exploit operational efficiencies through centralization of baseband functions, pooling efficiencies for RAN baseband processing, and air-interface performance gains by fast-time-scale multi-cell coordination. In this paper, we aim to derive insights underlying key technologies and tradeoffs that drive Cloud RAN. We highlight principles that explain the impact of Cloud RAN architectures and functional splits on multi-cell coordination gains. We show that the ability to extract pooling gains can differ significantly between RAN functions, depending on whether they are per-user or per-cell functions. We propose a method for elastic scaling of RAN functions that takes into account both real-time and non-real-time needs. We conclude with a proposal for a methodology for operators to evaluate the overall tradeoffs they will face in deciding whether to adopt Cloud RAN.

Centralized RAN and baseband pooling architectures are increasingly being seen as the way cellular technologies will be deployed in the future due to a number of advantages they provide to the operator. However, LTE and LTE-Advanced were designed with a distributed architecture in mind. Therefore, there are certain inherent technical issues when the link connecting the baseband and RRU (Remote Radio Unit) has a non-negligible latency. In this paper, we study the impact of this non-ideal fronthaul on the downlink and uplink performance of LTE-Advanced networks. We show that, while the impact on the peak throughput is quite severe, it is more modest when we have a larger number of active UEs. The degradation in performance is larger in the uplink than in the downlink due to the synchronous HARQ mechanism used in LTE uplink. In the downlink, we compare the performance of centralized RAN with that of distributed RAN in case of non-ideal fronthaul (FH) and backhaul (BH), respectively, and determine whether it is better to incur backhaul or fronthaul latency from a performance perspective. In the uplink, we analyze the loss in performance due to fronthaul latency. We have developed a novel look-ahead scheduling scheme that results in significant improvement in performance, especially when the number of active UEs is small.

In this big data era, huge amount data are continuously acquired for a variety of purposes. Advanced computing, imaging, and sensing technologies enable scientists to study natural and physical phenomena at unprecedented precision, resulting in an explosive growth of data. It is a huge challenge to visualize this growing data in static or in dynamic form. Most traditional data visualization approaches and tools can't support at "big" scale. In this paper, we identified the challenges and opportunities in big data visualization and review some current approaches and visualization tools.

Cloud computing is a promising and expanding technology which could replace traditional IT systems. Cloud computing resembles a giant pool of resources which contains hardware, software and related applications, which can be accessed through web-based services on a pay-per-usage model. The main features of the cloud model are accessibility, availability and scalability, and it can be subdivided into three service models: Software as a Service (SaaS), Platform as a Service (PaaS) and Infrastructure as a Service (IaaS). Cloud computing continues to transform how security challenges are addressed in closed and private networks. Given the advanced functionality offered by cloud computing, network monitoring and digital forensics efforts are potentially detectable and service-interruptive, which significantly impacts the effectiveness and thoroughness of digital forensic methods. This paper presents a general view of cloud computing, which aims to highlight the security issues and vulnerabilities associated with cloud service models. The technology is mainly based on virtualization, where data is always volatile and typically stored in a de-centralized architecture located across various countries and regions. This presents forensics investigators with legal challenges, due to the nature of multi-tenancy and distributed shared resources. This paper examines the three cloud service models and discusses the security challenges and issues involved with each service model along with potential solutions for each.

Multi-Cell coordination or Coordinated Multi-Point (CoMP) schemes are seen as the next evolution of LTE-Advanced to yield a further improvement in overall system performance. In this paper, we study in detail the performance of one such CoMP scheme, namely Multi-cell coordinated scheduling (MuCCS). In our previous work, we have shown that the performance of MuCCS using a distributed "liquid cooperation cluster," where each cell coordinates with its nearest neighbors, out-performs a centralized scheme where the coordination is done over a small set of non-overlapping cells (either three cell or nine cells). In this work, we answer the question of whether there are any additional gains to be obtained by a centralized implementation, if the cooperation cluster is allowed to be arbitrarily large. We also quantify the additional performance gain obtained from allowing intermediate transmit power values besides just the binary on/off power values. Finally, we evaluate the performance of MuCCS when the eNodeBs are connected by a non-ideal backhaul, and devise a distributed MuCCS scheme that retains its advantage over a centralized coordination scheme even in the presence of a non-ideal backhaul. Overall, our results validate that additional high-complexity extensions of coordinated scheduling does not yield large gains over the distributed MuCCS scheme with "liquid cooperation clusters" and on/off power values. In this paper, we use the geometric mean of UE throughputs as the performance metric, and we motivate the rationale behind use of this as a single metric for performance comparison.

In this paper, we consider the problem of uplink coordinated multi-point (UL CoMP) in a wireless network with non-ideal backhaul links. UL CoMP with joint receivers has the potential to improve both cell-edge and system capacity depending on the deployment environment. Realizing the performance gains from CoMP requires ideally large bandwidth and low-latency backhaul links to do data exchange. Even if backhaul bandwidth is well-provisioned, latencies in the backhaul data exchange protocols could lead to severe performance degradation due to the requirement of strict Ack/Nack timing imposed by the LTE standard [1]. In order to deal with this issue, we propose strategies for adaptive helper cell management in order to entail minimal performance loss across different loading conditions. Using system simulations and detailed performance analysis, we show the advantage of using the proposed approaches with a finite buffer traffic model for a heterogeneous network deployment scenario including a Macro-cell layer and a clustered small-cell layer.

Managing inter-cell interference is one of the main challenges in current and next generation wireless systems that aggressively reuse the frequency. Cooperation between interfering cells has been sought to mitigate interference. In this paper, we address the problem of jointly optimizing the transmit powers, user scheduling, and user association in a cellular network to maximize the weighted sum rate (WSR). To this end, we develop a distributed interference penalty algorithm in which the cells update their transmit powers and user schedule to maximize its utility minus an interference cost. The proposed algorithm involves only limited exchange of information via backhaul and has convergence guarantees. Furthermore, we propose a sub-optimal algorithm with lower computational and backhaul overhead. In it, the users are first associated to the base stations (BSs) based on their signal-to-interference-plus-noise-ratios (SINRs). It is then followed by joint optimization of BS transmit powers and user scheduling, for which we develop an interference penalty algorithm as well. We show that the proposed algorithms outperform the computationally complex weighted minimum mean squared error (WMMSE) algorithm.

Big data is a new technology wave that makes the world awash in data. Various organizations accumulate data that are difficult to exploit. Government databases, social media, healthcare databases etc. are the examples of the big data. Big data covers absorbing and analyzing huge amount of data that may have originated or processed outside of the organization. Data provenance can be defined as origin and process of data. It carries significant information of a system. It can be useful for debugging, auditing, measuring performance and trust in data. Data provenance in big data is relatively unexplored topic. It is necessary to appropriately track the creation and collection process of the data to provide context and reproducibility. In this paper, we propose an intuitive layer based architecture of data provenance and visualization. In addition, we show a complete workflow of tracking provenance information of big data.

Big data is a new technology wave that makes the world awash in data. Various organizations accumulate data that are difficult to exploit. Government databases, social media, healthcare databases etc. are the examples of that big data. Big data covers absorbing and analyzing huge amount of data that may have originated or processed outside of the organization. Data provenance can be defined as origin and process of data. It carries significant information of a system. It can be useful for debugging, auditing, measuring performance and trust in data. Data provenance in big data is relatively unexplored topic. It is necessary to appropriately track the creation and collection process of the data to provide context and reproducibility. This poster tries to address the challenges of capturing provenance data. Additionally, we propose an intuitive layer based architecture of provenance in big data that can handle the challenges.

The enormous volumes of data created and maintained by industries, research institutions are on the verge of outgrowing its infrastructure. The advancements in the organization's work flow include data storage, data management, data maintenance, data integration, and data interoperability. Among these levels, data integration and data interoperability can be the two major focus areas for the organizations which tend to implement advancements in their workflow. Overall, data integration and data interoperability influence the organization's performance. The data integration and data interoperability are complex challenges for the organizations deploying big data architectures due to the heterogeneous nature of data used by them. Therefore, it requires a comprehensive approach to negotiate the challenges in integration and interoperability. This paper focuses on the challenges of data integration and data interoperability in big data.

We compare the algorithms for centralized and decentralized multi-cell joint optimization of coordinated scheduling with muting. For conventional non-overlapping clusters of cooperating cells, we show that for the same latency of communication, an appropriate decentralized algorithm can equivalently achieve the same jointly optimal solution achieved by a centralized approach. In this decentralized algorithm, cells do not exchange CSIs of all users with other cells, but instead exchange per-cell benefit metrics. In non-overlapping clusters, border cells in a cluster do not coordinate with strong interferers outside the cluster. We introduce a new cluster design called Liquid Cluster which allows a cell to flexibly cooperate with all its relevant interferers. The decentralized algorithm easily extends to Liquid Clusters, performing better than conventional non-overlapping 9-cell clusters.

Heterogeneous networks (Hetnets) introduce an imbalance between downlink and uplink because the downlink transmit powers of the macro and pico eNodeBs can differ by as much as 16 dB. However, the user equipment (UE)'s maximum uplink transmit power is the same regardless of whether the UE is connected to the macro or the pico eNodeB. The uplink throughput performance depends both on the SINR that the UEs see and the number of PRBs that each UE can get allocated. Due to the difference in the transmit powers of the eNodeBs, the number of UEs that attach to the pico eNodeB is typically smaller than the number that attach to the macro eNodeB. In this work, we consider these impacts and optimize the power control settings of the macro and pico UEs so that the cell-edge and average UE throughput are improved. We study the performance under two different power control schemes: Fractional Power Control (FPC) and Interference Penalty Algorithm (IPA). Using simulations, we study the impact of various parameter settings on the performance of IPA. Based on our studies, we can conclude that the macro UE's transmit power needs to be larger than that of the pico UE to overcome the larger number of UEs attaching to the macro eNodeB. We also study the impact of imperfect knowledge of neighbor cell path losses on the uplink performance of IPA.

Dynamic Point selection (DPS) is a key Downlink (DL) Coordinated Multipoint (CoMP) technique that switches the serving data Transmission Point (TP) of a User Equipment (UE) dynamically among the UE's cooperating set of TPs without requiring a cell handover. It provides performance improvement due to TP selection-diversity gains and dynamic UE load balancing benefit. In this paper, we propose two simple DPS schemes that take into account the UE's current channel conditions and the cell loading conditions to make the UE's TP switching decisions. We show that these schemes improve the system performance under different practical and realistic settings, such as, cell handover margin, TP switching periods, bursty traffic conditions, and cooperation cell cluster sizes.

The 21st century has seen an enormous and almost sudden expansion in the use and types of technology. However, the cyber-age has also brought along with it new ways to wage attacks which has been a daunting task to thwart. This paper analyzes the number of high-impact security threats, vulnerabilities and alerts that have been reported by the United States Computer Emergency Readiness Team (US-CERT) over the past five years. This paper also closely examines the companies with the highest numbers of reports.

Studies related to public transportation systems help the commuting public by increasing road safety and circulation. These result in optimized traffic flow, shorter origin-destination travel time and reduced incident rate. Vehicular Ad-Hoc Network uses a number of sensors to gather data on the road. Intelligent Transportation Systems (ITS) draw inference from the gathered data. In this paper we discuss our experience of using Auto Regressive Integrated Moving Average (ARIMA) based techniques emphasizing on the integration of short-range and long-range dependencies of the historical traffic volume. We also analyze traffic data for patterns across different types of roads and derive computational complexity of ARIMA. Finally, improvements are identified for better prediction. We empirically show that SARIMA and ARIMA-GARCH exhibit similar road traffic prediction. ARIMA-GARCH is better than ARIMA and SARIMA for prediction, with stable model order across different historical traffic volumes. We further analyzes model orders across different types of roads and historical traffic volume; and its implications for practical applicability in ITS.

Utilizing data mining techniques to detect health care fraud has helped detect fraud patterns by providers and reduce the amount of waste spent and abuse in the health care system. However, less attention has been paid on Medicaid eligibility frauds by applicants in the health care system. In this paper, we explore fraud detection using a data driven approach at the time of eligibility application process. Identifying fraud at the beginning phase reduces the number of fraud abusers in the system and allow for future monitoring for such similar activities. We propose a generic data driven approach, which utilizes major public and private databases for detecting fraud and inaccurate eligibility requirements.

Multi-User MIMO (MU-MIMO) is a useful technique to enhance uplink capacity in LTE because the base stations employ multiple receive antennas whereas the user equipments transmit with only a single antenna. In this paper we investigate several UE pairing schemes that maximize the total system utility while adhering to the LTE Release-8 constraint of allocating identical contiguous resource blocks (RBs) to the paired UEs. We design pairing decisions using either the wideband or sub- band channel cross-correlation values of the UEs. These mechanisms offer varying trade-offs between performance and complexity.We show that with appropriate modifications, these schemes work well even for finite queue traffic.

In this paper, we develop a novel inter-cell interference co-ordination scheme that takes into account the interference cost on neighboring cells. We formulate a multi-cell utility maximization problem and subsequently decouple it into single-cell optimization problems by including an interference penalty. By solving this decoupled problem, we derive policies for user selection, resource allocation, and power-control. Since the coupling between cells is indirectly taken into account by means of the user's channel gain to the neighboring cells, our simulation results show that this distributed solution has no degradation in performance while little or no inter-cell co-ordination is required. We present simulation results that show that the Interference Penalty Algorithm (IPA) provides significant improvement in sector and cell-edge throughputs.

In this paper, the scheduling and resource allocation problem for the downlink in a code-division multiple access (CDMA)-based wireless network is considered. The problem is to select a subset of the users for transmission and for each of the users selected, to choose the modulation and coding scheme, transmission power, and number of codes used. We refer to this combination as the physical layer operating point (PLOP). Each PLOP consumes different amounts of code and power resources. The resource allocation task is to pick the ¿optimal¿ PLOP taking into account both system-wide and individual user resource constraints that can arise in a practical system. This problem is tackled as part of a utility maximization problem framed in earlier papers that includes both scheduling and resource allocation. In this setting, the problem reduces to maximizing the weighted throughput over the state-dependent downlink capacity region while taking into account the system-wide and individual user constraints. This problem is studied for the downlink of a Gaussian broadcast channel with orthogonal CDMA transmissions. This results in a tractable convex optimization problem. A dual formulation is used to obtain several key structural properties. By exploiting this structure, algorithms are developed to find the optimal solution with geometric convergence.

In this paper, we consider the single frequency network (SFN) operation of broadcast/multicast services, and develop resource allocation schemes to minimize the amount of resource over-provisioning. Over-provisioning is required when SFN areas are overlapping because all the cells in the SFN area should use the same resource to transmit a given MBMS service. We first develop an optimal resource allocation scheme that minimizes the over-provisioning. Due to the computational complexity of this scheme, we develop sub-optimal schemes whose complexity is quadratic in the number of services. In addition, we develop an incremental resource allocation scheme for newly arriving services that does not alter the resources allocated to existing services. We present numerical results showing the benefits of our proposed schemes. Motivated by these numerical results, we design a hierarchically overlapping SFN areas scheme that strikes the right balance between complexity and resource over-provisioning.

Orthogonal frequency division multiplexing (OFDM) with dynamic scheduling and resource allocation is a key component of most emerging broadband wireless access networks such as WiMAX and LTE (long term evolution) for 3GPP. However, scheduling and resource allocation in an OFDM system is complicated, especially in the uplink due to two reasons: (i) the discrete nature of subchannel assignments, and (ii) the heterogeneity of the users' subchannel conditions, individual resource constraints and application requirements. We approach this problem using a gradient-based scheduling framework. Physical layer resources (bandwidth and power) are allocated to maximize the projection onto the gradient of a total system utility function which models application-layer Quality of Service (QoS). This is formulated as a convex optimization problem and solved using a dual decomposition approach. This optimal solution has prohibitively high computational complexity but reveals guiding principles that we use to generate lower complexity sub-optimal algorithms. We analyze the complexity and compare the performance of these algorithms via extensive simulations.

We consider scheduling and resource allocation for the downlink of a cellular OFDM system, with various practical considerations including integer tone allocations, different sub-channelization schemes, maximum SNR constraint per tone, and "self-noise" due to channel estimation errors and phase noise. During each time-slot a subset of users must be scheduled, and the available tones and transmission power must be allocated among them. Employing a gradient-based scheduling scheme presented in earlier papers reduces this to an optimization problem to be solved in each time-slot. Using a dual formulation, we give an optimal algorithm for this problem when multiple users can time-share each tone. We then give several low complexity heuristics that enforce integer tone allocations. Simulations are used to compare the performance of different algorithms.

As 3G wireless systems evolve towards 4G, various wireless network technology organizations looking at network architectures for 4G are considering a redesign of the network away from the traditional centralized, hierarchical design towards a more distributed operation of network functions. In this paper, we present some mechanisms for distributed operation of paging in 4G systems. Our focus is on highlighting the tradeoffs that drive the design choice of whether to centralize or distribute the paging operation. One such tradeoff arises from considerations of service reliability and availability. We present a distributed redundancy scheme for improving the reliability and availability for distributed paging. We analyze the fraction of time that mobiles remain reachable through paging (i.e. the paging service availability) in idle mode in the presence of base station and backhaul failures.

Abstract For studying performance characteristics of radio channels, the knowledge about the probability density function (pdf) of fading–shadowing effects is essential. K‐distribution corresponding to Rayleigh–gamma distribution (RGD) is widely used to approximate a more realistic Rayleigh–lognormal distribution (RLD) which does not have a closed form expression. A new composite Rayleigh‐inverse Gaussian distribution (RIGD), an alternative to K‐distribution, is analyzed with regards to its suitability and effectiveness in radio channels. Detailed investigations are made to study the performance characteristics of RIGD and K‐distribution (RGD) in terms of Kullback–Leibler (KL) measure of divergence. Based on these investigations, it is found that RIGD is better suited for capturing fading–shadowing aspects of radio channels instead of K‐distribution. Copyright © 2006 John Wiley & Sons, Ltd.

In this paper, we propose a technique to retrieve the images using the 'search by similarity' method with the help of multimodal keywords. Multimodal keywords consist of low-level MPEG-7 color descriptors and textual keywords. The visual keywords and textual keywords are combined together and the image collection is represented as a matrix, which is similar in representation to a term-document matrix. Using LSI (latent semantic indexing), we demonstrate that the visual keywords, when combined with textual keywords can improve the retrieval results to a great degree

In a GPRS or EDGE system, fluctuations in the wireless link rate can cause the buffer in the BSS to either overflow or underflow if the data flow into the BSS buffer from the SGSN is not controlled properly. Currently available algorithms for controlling this data flow rely on imperfect estimates of the wireless link rate and cannot ensure that the BSS buffer will never overflow or underflow. We present a new low-complexity algorithm that ensures that the BSS buffer will never overflow or underflow, without requiring any estimate of the wireless link rate. If there is no feedback delay between the SGSN and BSS, the algorithm is optimal: among all algorithms that ensure that the BSS buffer will never overflow or underflow, it requires the smallest amount of buffering in the BSS. The algorithm also has a fixed point, i.e. if the wireless link rate remains constant, the BSS buffer remains in equilibrium and no flow control updates are required.

We discuss a simple conceptual framework for analyzing the flow of data in integrated services networks. The framework allows us to easily model and analyze the behavior of open loop, rate based flow control protocols, as well as closed loop, window based flow control protocols. Central to the framework is the concept of a service curve element, whose departure process is bounded between the convolution of the arrival process with a minimum service curve and the convolution of the arrival process with a maximum service curve. Service curve elements can model links, propagation delays, schedulers, regulators, and window based throttles. The mathematical properties of convolution allow us to easily analyze complex configurations of service curve elements to obtain bounds on the end-to-end performance. We demonstrate this by examples, and investigate tradeoffs between buffering requirements, throughput, and delay, for different flow control strategies.

We consider a model comprising of a single bottleneck node/switch fed by multiple Markov modulated fluid sources and with significant and possibly different propagation delays. The amount of traffic released by the sources depends on its state as well as delayed feedback from the network. Such a source model allows us to capture ABR as well as VBR traffic types. We obtain analytical results for computing various performance measures associated with this model. Based on these analytical results, we study the performance of the feedback policy for two different delay classes. In the parameter regime investigated, we observe that while the sources with a smaller delay do better, they do not do significantly better. Moreover, the performance of the farther sources improves as the second set of sources is moved closer. Thus, what may be perceived as unfairness due to the difference in feedback delays should really be interpreted as the closer sources helping the farther sources while helping themselves.

We consider a system comprising of a single bottleneck switch/node that is fed by N independent Markov-modulated fluid sources. There is a fixed propagation delay incurred by the traffic between these sources and the switch. We assume that the switch sends periodic feedback in the form of a single congestion indicator bit. This feedback also incurs a fixed propagation delay in reaching the sources. Upon reaching the sources (or the access controllers associated with the sources), this congestion indicator bit is used to choose between two rates for the excess traffic, high or low, possibly depending on the state of that source. The switch employs a threshold mechanism based on its buffer level to discard excess traffic. We show that the stationary distribution of this system satisfies a set of first-order linear differential equations along with a set of split boundary conditions. We obtain an explicit solution to these using spectral decomposition. To this end we investigate the related eigenvalue problem. Based on these results we investigate the role of delayed feedback vis-a-vis various time-constants and traffic parameters associated with the system. In particular, we identify conditions under which the feedback scheme offers significant improvement over the open-loop scheme.>