In this paper, we propose a Neural Radiance Fields (NeRF) based framework, referred to as Novel View Synthesis Framework (NVSF). It jointly learns the implicit neural representation of space and time-varying scene for both LiDAR and Camera. We test this on a real-world autonomous driving scenario containing both static and dynamic scenes. Compared to existing multimodal dynamic NeRFs, our framework is self-supervised, thus eliminating the need for 3D labels. For efficient training and faster convergence, we introduce heuristic-based image pixel sampling to focus on pixels with rich information. To preserve the local features of LiDAR points, a Double Gradient based mask is employed. Extensive experiments on the KITTI-360 dataset show that, compared to the baseline models, our framework has reported best performance on both LiDAR and Camera domain. Code of the model is available at https://github.com/gaurav00700/Selfsupervised-NVSF

Radar is an inevitable part of the perception sensor set for autonomous driving functions. It plays a gap-filling role to complement the shortcomings of other sensors in diverse scenarios and weather conditions.In this paper, we propose a Deep Neural Network (DNN) based end-to-end object detection and heading estimation framework using raw radar data. To this end, we approach the problem in both a "Data-centric" and "model-centric" manner. We refine the publicly available CARRADA [1] dataset and introduce Bivariate norm annotations. Besides, the baseline model is improved by a transformer [2] inspired cross-attention fusion and further center-offset maps are added to reduce localisation error. Our proposed model improves the detection mean Average Precision (mAP) by 5%, while reducing the model complexity by almost 23%. For comprehensive scene understanding purposes, we extend our model for heading estimation. The improved ground truth and proposed model is available at Github.

Increasingly sophisticated image editing tools make it easy to modify images. Often these modifications are elaborate, convincing, and undetectable by even careful human inspection. These considerations have prompted the development of forensic algorithms and approaches to detect modifications done to an image. However, these detectors are model-driven (i.e., manipulation-specific) and the choice of a potent detector requires knowledge of the type of manipulation, something that cannot be known ( a priori ). Thus, the latest effort is directed towards developing model-free (i.e., generalized) detectors capable of detecting multiple manipulation types. In this article, we propose a novel detector capable of exposing seven different manipulation types in low-resolution compressed images. Our proposed approach is based on a two-layer convolutional neural network (CNN) to extract frequency domain features of image median filtered residual that are classified using two different classifiers—softmax and extremely randomized trees. Extensive experiments demonstrate the efficacy of proposed detector over existing state-of-the-art detectors.

In this paper, we shed light on how powerful congestion control based on local interactions may be obtained. We show how ants can use repellent pheromones and incorporate the effect of crowding to avoid traffic congestion on the optimal path. Based on these interactions, we propose an ant algorithm, the M-unit EigenAnt algorithm, that leads to the selection of the M shortest paths. The ratio of selection of each of these paths is also optimal and regulated by an optimal amount of pheromone on each of them. To the best of our knowledge, the M-unit EigenAnt algorithm is the first antalgorithm that explicitly ensures the selection of the M shortest paths and regulates the amount of pheromone on them such that it is asymptotically optimal. In fact, it is in contrast with most ant algorithms that aim to discover just a single best path. We provide its convergence analysis and show that the steady state distribution of pheromone aligns with the eigenvectors of the cost matrix, and thus is related to its measure of quality. We also provide analysis to show that this property ensues even when the food is moved or path lengths change during foraging. We show that this behavior is robust in the presence of fluctuations and quickly reflects the change in the M optimal solutions. This makes it suitable for not only distributed applications butalso dynamic ones as well. Finally, we provide simulation results for the convergence to the optimal solution under different initial biases, dynamism in lengths of paths, and discovery of new paths.

Various studies have pointed out the debilitating effects of OS jitter on the performance of parallel applications on large clusters such as the ASCI Purple and the Mare Nostrum at Barcelona Supercomputing Center. These clusters use commodity OSes such as AIX and Linux respectively. The biggest hindrance in evaluating any technique to mitigate jitter is getting access to such large scale production HPC systems running a commodity OS. An earlier attempt aimed at solving this problem was to emulate the effects of OS jitter on more widely available and jitter-free systems such as BlueGene/L. In this paper, we point out the shortcomings of previous such approaches and present the design and implementation of an emulation framework that helps overcome those shortcomings by using innovative techniques. We collect jitter traces on a commodity OS with a given configuration, under which we want to study the scaling behavior. These traces are then replayed on a jitter-free system to predict scalability in presence of OS jitter. The application of this emulation framework to predict scalability is illustrated through a comparative scalability study of an off-the-shelf Linux distribution with a minimal configuration (runlevel 1) and a highly optimized embedded Linux distribution, running on the IO nodes of BlueGene/L. We validate the results of our emulation both on a single node as well as on a real cluster. Our results indicate that an optimized OS along with a technique to synchronize jitter can reduce the performance degradation due to jitter from 99% (in case of the off-the-shelf Linux without any synchronization) to a much more tolerable level of 6% (in case of highly optimized BlueGene/L IO node Linux with synchronization) at 2048 processors. Furthermore, perfect synchronization can give linear scaling with less than 1% slowdown, regardless of the type of OS used. However, as the jitter at different nodes starts getting desynchronized, even with a minor skew across nodes, the optimized OS starts outperforming the off-the-shelf OS.

Projection of high dimensional data into a lower dimensional subspace is required for human understanding of the health of an IT infrastructure. Over the past several years, there have been a large number of dimensionality reduction techniques that have been proposed. Their direct application for visualizing system monitoring data is challenged by two factors. First, system monitoring data does not lie in a metric space. Second, system monitoring data is intrinsically iquestmulti-resolutioniquest in that an event may lead to cascaded events (a server going down impacts one or more applications running on the server; a network outage may impact several network dependent components). Lower dimensional representations which do not take into account the intrinsic multi-resolution nature of the monitoring data are thus limited in their utility and challenge human comprehension. In this paper, we exploit the multi-resolution nature of the monitoring data and a 1-of-n representation of event data to construct navigable multi-resolution topology preserving views for visualizing system monitoring data. We also demonstrate the efficacy of the proposed approach using data from a real data center.

Understanding the behavior and impact of various sources of Operating System Jitter (OS Jitter) is important not only for tuning a system for HPC applications, but also for the ongoing efforts to create light-weight versions of commercial operating systems such as Linux, that can be used on compute nodes of large scale HPC systems. In this paper, we present a tool that helps in identifying sources of OS Jitter in a commodity operating system such as Linux and measures the impact of OS Jitter through fine grained kernel instrumentation. Our methodology comprises of running a user-level micro-benchmark and measuring the latencies experienced by the benchmark. We then associate each latency to operating system daemons and interrupts using data obtained from kernel instrumentation. We present experimental results that help identify the biggest contributors to the total OS Jitter perceived by an application on a commodity operating system such as Linux. Our results revealed that while 63% of the total jitter comes from timer interrupts, the rest comes from various system daemons and interrupts, most of which can be easily eliminated. The tool presented in this paper can also be used to tune “out of the box” commodity operating systems as well as to detect new sources of operating system jitter that get introduced as software get installed and upgraded on a tuned system.

Supervised learning algorithms (relevance feedback (RF) algorithms) are often used in content based image retrieval (CBIR) systems to enhance interactive search and browsing of image databases. One of the issues associated with RF based CBIR systems is the lack of a large training set. Labeling of images is a time consuming activity and users usually do not have the patience to provide feedback on a large set. Thus the challenge is to select a "good" small training set in order to improve the retrieval performance of CBIR systems. In this paper we propose to use orthogonal arrays (OA), popularly used in design of experiments, in order to select this set. The property of OA that make them useful for CBIR systems is that they can be designed with or without using any prior classification information. We show that the top k retrieval accuracy increases rapidly as the number of labeled samples increase.

Top-down induction of decision trees is a simple and powerful method of pattern classification. In a decision tree, each node partitions the available patterns into two or more sets. New nodes are created to handle each of the resulting partitions and the process continues. A node is considered terminal if it satisfies some stopping criteria (for example, purity, i.e., all patterns at the node are from a single class). Decision trees may be univariate, linear multivariate, or nonlinear multivariate depending on whether a single attribute, a linear function of all the attributes, or a nonlinear function of all the attributes is used for the partitioning at each node of the decision tree. Though nonlinear multivariate decision trees are the most powerful, they are more susceptible to the risks of overfitting. In this paper, we propose to perform model selection at each decision node to build omnivariate decision trees. The model selection is done using a novel classifiability measure that captures the possible sources of misclassification with relative ease and is able to accurately reflect the complexity of the subproblem at each node. The proposed approach is fast and does not suffer from as high a computational burden as that incurred by typical model selection algorithms. Empirical results over 26 data sets indicate that our approach is faster and achieves better classification accuracy compared to statistical model select algorithms.

In general, pattern classification algorithms assume that all the features are available during the construction of a classifier and its subsequent use. In many practical situations, data are recorded in different servers that are geographically apart, and each server observes features of local interest. The underlying infrastructure and other logistics (such as access control) in many cases do not permit continual synchronization. Each server thus has a partial view of the data in the sense that feature subsets (not necessarily disjoint) are available at each server. In this article, we present a classification algorithm for this distributed vertically partitioned data. We assume that local classifiers can be constructed based on the local partial views of the data available at each server. These local classifiers can be any one of the many standard classifiers (e.g., neural networks, decision tree, k nearest neighbor). Often these local classifiers are constructed to support decision making at each location, and our focus is not on these individual local classifiers. Rather, our focus is constructing a classifier that can use these local classifiers to achieve an error rate that is as close as possible to that of a classifier having access to the entire feature set. We empirically demonstrate the efficacy of the proposed algorithm and also provide theoretical results quantifying the loss that results as compared to the situation where the entire feature set is available to any single classifier.

Journals and conference proceedings represent the dominant mechanisms of reporting new biomedical results. The unstructured nature of such publications makes it difficult to utilize data mining or automated knowledge discovery techniques. Annotation (or markup) of these unstructured documents represents the first step in making these documents machine analyzable. In this paper we first present a system called BioAnnotator for identifying and annotating biological terms in documents. BioAnnotator uses domain based dictionary look-up for recognizing known terms and a rule engine for discovering new terms. The combination and dictionary look-up and rules result in good performance (87% precision and 94% recall on the GENIA 1.1 corpus for extracting general biological terms based on an approximate matching criterion). To demonstrate the subsequent mining and knowledge discovery activities that are made feasible by BioAnnotator, we also present a system called MedSummarizer that uses the extracted terms to identify the common concepts in a given group of genes.

Supervised learning algorithms (relevance feedback (RF) algorithms) are often used in content based image retrieval (CBIR) systems to enhance interactive search and browsing of image databases. One of the issues associated with RF based CBIR systems is the lack of a large training set. Labeling of images is a time consuming activity and user's usually do not have the patience to label a large set. The challenge is to somehow leverage the much larger set of unlabeled images to improve the performance of CBIR systems. In this paper we propose a novel RF algorithm which learns from both labeled and unlabeled data. Our proposed algorithm also uses active learning so as to maximize the information gained from a given amount of user feedback.

Clickstream analysis provides valuable insight into the behavior of users and can be translated into better business opportunities and increased user satisfaction. A fundamental problem in clickstream analysis is the computation of the distance (or the similarity) between two clickstreams. While, there exists a considerable amount of literature which propose methods of computing path similarities, they rely on the edit distance or the related longest common subsequence to align the two clickstreams. The edit distance provides a least cost sequence of transformations that result in the two clickstreams to be identical. Often, measures of path similarity are defined on these “aligned” clickstreams. However, the replacement cost used in the “alignment” process used by the edit distance is assumed to be fixed and ignores the degree of similarity of the two page views. Proposed in this paper is a method for computing the replacement cost that is based on the assumption that the degree of similarity between two page views is proportional to their relative frequency of cooccurrence. We define a method, which includes the order of the sequence as well as the time spent on each page, for obtaining the replacement cost of two arbitrary web pages. Though less accurate than content based analysis, our experiments with data generated from a simulator as well as data from an actual web site show that our assumption is well founded and that the proposed method provides a fast and accurate method of computing the similarity between two page views.

This study explores the use of artificial neural networks (ANNs) models and brightness temperature from the Southern Great Plains in the United States to classify soil into different textures. Previous studies using ANN models and brightness temperature in a single drying cycle suggested that they might contain sufficient features to classify soil into three categories. To classify soil into more than three groups and to explore the limits of classification accuracy, this paper suggests the use of multiple-drying-cycle brightness temperature data. We have performed several experiments with feedforward neural network (FFNN) models, and the results suggest that the maximum achievable classification accuracy through the use of multiple-drying-cycle brightness temperature is about 80%. It appears that the rapidly changing space-time evolution of brightness temperature will restrict the FFNN model performance. Motivated by these observations, we have used a simple prototype-based classifier, known as the 1-NN model, and achieved 86% classification accuracy for six textural groups. A comparison of error regions predicted by both models suggests that for the given input representation maximum achievable accuracy for classification into six soil texture types is about 93%.

The considerable time and expense required for labeling data has prompted the development of algorithms which maximize the classification accuracy for a given amount of labeling effort. On the one hand, the effort has been to develop the so-called "active learning" algorithms which sequentially choose the patterns to be explicitly labeled so as to realize the maximum information gain from each labeling. On the other hand, the effort has been to develop algorithms that can learn from labeled as well as the more abundant unlabeled data. Proposed in this paper is an algorithm that integrates the benefits of active learning with the benefits of learning from labeled and unlabeled data. Our approach is based on reversing the roles of the labeled and unlabeled data. Specifically, we use a Genetic Algorithm (GA) to iteratively refine the class membership of the unlabeled patterns so that the maximum a posteriori (MAP) based predicted labels of the patterns in the labeled dataset are in agreement with the known labels. This reversal of the role of labeled and unlabeled patterns leads to an implicit class assignment of the unlabeled patterns. For active learning, we use a subset of the GA population to construct multiple MAP classifiers. Points in the input space where there is maximal disagreement amongst these classifiers are then selected for explicit labeling. The learning from labeled and unlabeled data and active learning phases are interlaced and together provide accurate classification while minimizing the labeling effort.

Decision tree induction is typically based on a top-down greedy algorithm that makes locally optimal decisions at each node. Due to the greedy and local nature of the decisions made at each node, there is considerable possibility of instances at the node being split along branches such that instances along some or all of the branches require a large number of additional nodes for classification. In this paper, we present a computationally efficient way of incorporating look-ahead into fuzzy decision tree induction. Our algorithm is based on establishing the decision at each internal node by jointly optimizing the node splitting criterion (information gain or gain ratio) and the classifiability of instances along each branch of the node. Simulations results confirm that the use of the proposed look-ahead method leads to smaller decision trees and as a consequence better test performance.

Linear projections for dimensionality reduction, computed using linear discriminant analysis (LDA), are commonly based on optimization of certain separability criteria in the output space. The resulting optimization problem is linear, but these separability criteria are not directly related to the classification accuracy in the output space. Consequently, a trial and error procedure has to be invoked, experimenting with different separability criteria that differ in the weighting function used and selecting the one that performed best on the training set. Often, even the best weighting function among the trial choices results in poor classification of data in the subspace. In this short paper, we introduce the concept of fractional dimensionality and develop an incremental procedure, called the fractional-step LDA (F-LDA) to reduce the dimensionality in fractional steps. The F-LDA algorithm is more robust to the selection of weighting function and for any given weighting function, it finds a subspace in which the classification accuracy is higher than that obtained using LDA.

Dimensionality reduction is the process of mapping high-dimension patterns to a lower dimension subspace. When done prior to classification, estimates obtained in the lower dimension subspace are more reliable. For some classifiers, there is also an improvement in performance due to the removal of the diluting effect of redundant information. A majority of the present approaches to dimensionality reduction are based on scatter matrices or other statistics of the data which do not directly correlate to classification accuracy. The optimality criteria of choice for the purposes of classification is the Bayes error. Usually however, Bayes error is difficult to express analytically. We propose an optimality criteria based on an approximation of the Bayes error and use it to formulate a linear and a nonlinear method of dimensionality reduction. The nonlinear method we propose, relies on using a multilayered perceptron which produces as output the lower dimensional representation. It thus differs from autoassociative like multilayered perceptrons which have been proposed and used for dimensionality reduction. Our results show that the nonlinear method is, as anticipated, superior to the linear method in that it can perform unfolding of a nonlinear manifold. In addition, the nonlinear method we propose provides substantially better lower dimension representation (for classification purposes) than Fisher's linear discriminant (FLD) and two other nonlinear methods of dimensionality reduction that are often used.

Subspace methods of pattern recognition form an interesting and popular classification paradigm. The earliest subspace method of classification was the class featuring information compression (CLAFIC) which associated with each class a linear subspace. Local subspace classification methodologies which have enhanced classification power by associating multiple linear subspaces with each class have also been investigated. In this paper, we introduce the oriented soft regional subspace classifier (OS-RSC). The highlights of this classifier are: (i) class specific subspaces are formed to specifically maximize the average projection of one class while minimizing that of the rival class; (ii) multiple manifolds are formed for each class increasing the classification power; and (iii) soft sharing of the training patterns again allows for consistent classification performance. It turns out that the cost function for forming class specific subspaces is maximized for a subspace of unit dimensionality. The performance of the proposed classifier is tested on real-world classification problems.

In this paper we introduce a new classifier, K-CLAFIC (Kernel based extension of CLAss Featuring Information Compression). CLAFIC is a statistical classification paradigm which associates with each output class a linear subspace. Thus patterns are classified based on their distance from different vector subspaces. Based on a newly introduced method to perform nonlinear principal component analysis, we present K-CLAFIC as a natural nonlinear extension of CLAFIC. Thus in K-CLAFIC there is a nonlinear subspace associated with each class and patterns are classified based on their distance from different nonlinear subspaces. K-CLAFIC is simple in operation and gives highly competitive performance on standard datasets. Also, since there are no iterative procedures for parameter optimization, in spite of being a nonlinear classifier, it is fast in operation.

Spatial interactions over a local neighborhood between neurons is common in many models of neural networks. The spatial interaction is often shaped to allow for improved response in the application for which the network is constructed. In this paper, we study an additional effect that can occur when such spatial interactions are imposed. Specifically, we find that the net input received by a neuron from neighboring neurons can induce stochastic resonance thereby increasing the signal-to-noise ratio of the neuron output. We also find that stochastic resonance is enhanced with increasing variance between the neighborhood signals received by the neuron. Thus the net signal received by a neuron due to neighborhood neurons plays a role similar to that played by noise in a classical stochastic resonance situations. We believe that these results, when combined with the more traditional use of spatial interaction, motivate the design of new and enhanced neural detectors.

Dimensionality reduction is the process of mapping high dimensional patterns to a lower dimensional manifold and is typically used for visualization or as a preprocessing step in classification applications. From a classification viewpoint, the rate of increase of Bayes error serves as an ideal choice to measure the loss of information relevant to classification. Motivated by that, we present a multilayer perceptron which produces as output the lower dimensional representation. The multilayer perceptron is trained so as to minimize the classification error in the subspace. It thus differs from autoassociative like multilayer perceptrons which have been proposed and used for dimensionality reduction. We examine the performance of the proposed method of dimensionality reduction and the effect that varying the parameters have on the algorithm.

Compact characterization of soil moisture at a given scale using self-organizing feature maps is presented. The authors find that as few as 49 neurons capture the spatial structure of remotely sensed soil moisture images from the southern Great Plains. Average latent heat flux computed from the original image of 21204 pixels and from 49 neurons are comparable.

In this article we study the effect of dynamically modifying the weight matrix on the performance of a neural associative memory. The dynamic modification is implemented by adding, at each step, the outer product of the current state, scaled by a suitable constant eta, to the correlation weight matrix. For single-shot synchronous dynamics, we analytically obtain the optimal value of eta. Although knowledge of the noise percentage is required for calculating the optimal value of eta, a fairly good choice of eta can be made even when the amount of noise is not known. Experimental results are provided in support of the analysis. The efficacy of the proposed modification is also experimentally verified for the case of asynchronous updating with transient length > 1.

Determining the optical flow is an ill-posed problem, and requires the inclusion of a regularization term for solution. We show that ordered maps produced through self-organization reflect the topological relationships of the input, and can thus inherently supply the constraints required in obtaining the optical flow. Our computational procedure is thus based on training a self-organizing feature map with features from the first frame of an image sequence, and observing the displacement weights in the weights when, the network is subsequently trained with features drawn from the second frame. We show through four simulations (three single object, and one multiple object) that the weight displacements provide an accurate representation of the optical flow.

We investigate the effects of including selected lateral interconnections in a feedforward neural network. In a network with one hidden layer consisting of m hidden neurons labeled 1,2… m, hidden neuron j is connected fully to the inputs, the outputs, and hidden neuron j + 1. As a consequence of the lateral connections, each hidden neuron receives two error signals: one from the output layer and one through the lateral interconnection. We show that the use of these lateral interconnections among the hidden-layer neurons facilitates controlled assignment of role and specialization of the hidden-layer neurons. In particular, we show that as training progresses, hidden neurons become progressively specialized—starting from the fringes (i.e., lower and higher numbered hidden neurons, e.g., 1, 2, m — 1 m) and leaving the neurons in the center of the hidden layer (i.e., hidden-layer neurons numbered close to m/2) unspecialized or functionally identical. Consequently, the network behaves like network growing algorithms without the explicit need to add hidden units, and like soft weight sharing due to functionally identical neurons in the center of the hidden layer. Experimental results from one classification and one function approximation problems are presented to illustrate selective specialization of the hidden-layer neurons. In addition, the improved generalization that results from a decrease in the effective number of free parameters is illustrated through a simple function approximation example and with a real-world data set. Besides the reduction in the number of free parameters, the localization of weight sharing may also allow for a method that allows procedural determination for the number of hidden-layer neurons required for a given learning task.

This paper describes a computational approach for accurately determining the location of human eyes in unconstrained monoscopic gray level images. The proposed method is based on exploiting the flow field characteristics that arise due to the presence of a dark iris surrounded by a light sclera. A novel aspect of the proposed method lies in its use of both spatial and temporal information to detect the location of the eyes. The spatial processing utilizes flow field information to select a pool of potential candidate locations for the eyes. Temporal processing uses the principle of continuity to filter out the actual location of the eyes from the pool of potential candidates. Extensions for gaze angle determination, and the tracking of human point-of-regard are indicated.