Deep neural networks with millions of parameters are at the heart of many state of the art computer vision models. However, recent works have shown that models with much smaller number of parameters can often perform just as well. A smaller model has the advantage of being faster to evaluate and easier to store - both of which are crucial for real-time and embedded applications. While prior work on compressing neural networks have looked at methods based on sparsity, quantization and factorization of neural network layers, we look at the alternate approach of pruning neurons. Training Neural Networks is often described as a kind of `black magic', as successful training requires setting the right hyper-parameter values (such as the number of neurons in a layer, depth of the network, etc ). It is often not clear what these values should be, and these decisions often end up being either ad-hoc or driven through extensive experimentation. It would be desirable to automatically set some of these hyper-parameters for the user so as to minimize trial-and-error. Combining this objective with our earlier preference for smaller models, we ask the following question - for a given task, is it possible to come up with small neural network architectures automatically? In this thesis, we propose methods to achieve the same. The work is divided into four parts. First, given a neural network, we look at the problem of identifying important and unimportant neurons. We look at this problem in a data-free setting, i.e; assuming that the data the neural network was trained on, is not available. We propose two rules for identifying wasteful neurons and show that these suffice in such a data-free setting. By removing neurons based on these rules, we are able to reduce model size without significantly affecting accuracy. Second, we propose an automated learning procedure to remove neurons during the process of training. We call this procedure ‘Architecture-Learning’, as this automatically discovers the optimal width and depth of neural networks. We empirically show that this procedure is preferable to trial-and-error based Bayesian Optimization procedures for selecting neural network architectures. Third, we connect ‘Architecture-Learning’ to a popular regularize called ‘Dropout’, and propose a novel regularized which we call ‘Generalized Dropout’. From a Bayesian viewpoint, this method corresponds to a hierarchical extension of the Dropout algorithm. Empirically, we observe that Generalized Dropout corresponds to a more flexible version of Dropout, and works in scenarios where Dropout fails. Finally, we apply our procedure for removing neurons to the problem of removing weights in a neural network, and achieve state-of-the-art results in scarifying neural networks.

Computational imaging has been playing an important role in understanding and analysing the captured images. Both image segmentation and restoration has been in-tegral parts of computational imaging. The studies performed in this thesis has been focussed toward developing novel algorithms for image segmentation and restoration. Study related to usage of Morozov Discrepancy Principle in Di use Optical Imaging was also presented here to show that hyper parameter selection could be performed with ease. The Laplacian of Gaussian (LoG) and Canny operators use Gaussian smoothing be-fore applying the derivative operator for edge detection in real images. The LoG kernel was based on second derivative and is highly sensitive to noise when compared to the Canny edge detector. A new edge detection kernel, called as Helmholtz of Gaussian (HoG), which provides higher di suavity is developed in this thesis and it was shown that it is more robust to noise. The formulation of the developed HoG kernel is similar to LoG. It was also shown both theoretically and experimentally that LoG is a special case of HoG. This kernel when used as an edge detector exhibited superior performance compared to LoG, Canny and wavelet based edge detector for the standard test cases both in one- and two-dimensions. The linear inverse problem encountered in restoration of blurred noisy images is typically solved via Tikhonov minimization. The outcome (restored image) of such min-imitation is highly dependent on the choice of regularization parameter. In the absence of prior information about the noise levels in the blurred image, ending this regular-inaction/hyper parameter in an automated way becomes extremely challenging. The available methods like Generalized Cross Validation (GCV) may not yield optimal re-salts in all cases. A novel method that relies on minimal residual method for ending the regularization parameter automatically was proposed here and was systematically compared with the GCV method. It was shown that the proposed method performance was superior to the GCV method in providing high quality restored images in cases where the noise levels are high Di use optical tomography uses near infrared (NIR) light as the probing media to recover the distributions of tissue optical properties with an ability to provide functional information of the tissue under investigation. As NIR light propagation in the tissue is dominated by scattering, the image reconstruction problem (inverse problem) is non-linear and ill-posed, requiring usage of advanced computational methods to compensate this. An automated method for selection of regularization/hyper parameter that incorporates Morozov discrepancy principle(MDP) into the Tikhonov method was proposed and shown to be a promising method for the dynamic Di use Optical Tomography.

Nowadays, we have millions of cameras in public places such as traffic junctions, railway stations etc., and capturing video data round the clock. This humongous data has resulted in an increased need for automation of visual surveillance. Analysis of crowd and traffic flows is an important step towards achieving this goal. In this work, we present our algorithms for identifying and segmenting dominant ows in surveillance scenarios. In the second part, we present our work aiming at predicting the visual saliency. The ability of humans to discriminate and selectively pay attention to few regions in the scene over the others is a key attentional mechanism. Here, we present our algorithms for predicting human eye fixations and segmenting salient objects in the scene. (i) Flow Analysis in Surveillance Videos: We propose algorithms for segmenting flows of static and dynamic nature in surveillance videos in an unsupervised manner. In static flows scenarios, we assume the motion patterns to be consistent over the entire duration of video and analyze them in the compressed domain using H.264 motion vectors. Our approach is based on modeling the motion vector field as a Conditional Random Field (CRF) and obtaining oriented motion segments which are merged to obtain the final flow segments. This approach in compressed domain is shown to be both accurate and computationally efficient. In the case of dynamic flow videos (e.g. flows at a traffic junction), we propose a method for segmenting the individual object flows over long durations. This long-term flow segmentation is achieved in the framework of CRF using local color and motion features. We propose a Dynamic Time Warping (DTW) based distance measure between flow segments for clustering them and generate representative dominant ow models. Using these dominant flow models, we perform path prediction for the vehicles entering the camera's field-of-view and detect anomalous motions. (ii) Visual Saliency Prediction using Deep Convolutional Neural Networks: We propose a deep fully convolutional neural network (CNN) - DeepFix, for accurately predicting eye fixations in the form of saliency maps. Unlike classical works which characterize the saliency map using various hand-crafted features, our model automatically learns features in a hierarchical fashion and predicts saliency map in an end-to-end manner. DeepFix is designed to capture visual semantics at multiple scales while taking global context into account. Generally, fully convolutional nets are spatially invariant which prevents them from modeling location dependent patterns (e.g. centre-bias). Our network overcomes this limitation by incorporating a novel Location Biased Convolutional layer. We experimentally show that our network outperforms other recent approaches by a significant margin. In general, human eye fixations correlate with locations of salient objects in the scene. However, only a handful of approaches have attempted to simultaneously address these related aspects of eye fixations and object saliency. In our work, we also propose a deep convolutional network capable of simultaneously predicting eye fixations and segmenting salient objects in a unified framework. We design the initial network layers, shared between both the tasks, such that they capture the global contextual aspects of saliency, while the deeper layers of the network address task specific aspects. Our network shows a significant improvement over the current state-of-the-art for both eye fixation prediction and salient object segmentation across a number of challenging datasets.

Enterprise Database Management Systems (DBMSs) have to contend with resource-intensive and time-varying workloads, making them well-suited candidates for migration to cloud plat-forms { specifically, they can dynamically leverage the resource elasticity while retaining affordability through the pay-as-you-go rental interface. The current design of database engine components lays emphasis on maximizing computing efficiency, but to fully capitalize on the cloud's benefits, the outlays of these computations also need to be factored into the planning exercise. In this thesis, we investigate this contemporary problem in the context of industrial-strength deployments of relational database systems on real-world cloud platforms. Specifically, we consider how the traditional metric used to compare query execution plans, namely response-time, can be augmented to incorporate monetary costs in the decision process. The challenge here is that execution-time and monetary costs are adversarial metrics, with a decrease in one entailing a rise in the other. For instance, a Virtual Machine (VM) with rich physical resources (RAM, cores, etc.) decreases the query response-time, but is expensive with regard to rental rates. In a nutshell, there is a tradeoff between money and time, and our goal therefore is to identify the VM that others the best tradeoff between these two competing considerations. In our study, we pro le the behavior of money versus time for a given query, and de ne the best tradeoff as the \knee" { that is, the location on the pro le with the minimum Euclidean distance from the origin. To study the performance of industrial-strength database engines on real-world cloud infrastructure, we have deployed a commercial DBMS on Google cloud services. On this platform, we have carried out extensive experimentation with the TPC-DS decision-support benchmark, an industry-wide standard for evaluating database system performance. Our experiments demonstrate that the choice of VM for hosting the database server is a crucial decision, because: (i) variation in time and money across VMs is significant for a given query, (ii) no one VM offers the best money-time tradeoff across all queries. To efficiently identify the VM with the best tradeoff from a large suite of available configurations, we propose a technique to characterize the money-time pro le for a given query. The core of this technique is a VM pruning mechanism that exploits the property of partially ordered set of the VMs on their resources. It processes the minimal and maximal VMs of this poset for estimated query response-time. If the response-times on these extreme VMs are similar, then all the VMs sandwiched between them are pruned from further consideration. Otherwise, the already processed VMs are set aside, and the minimal and maximal VMs of the remaining unprocessed VMs are evaluated for their response-times. Finally, the knee VM is identified from the processed VMs as the one with the minimum Euclidean distance from the origin on the money-time space. We theoretically prove that this technique always identifies the knee VM; further, if it is acceptable to and a \near-optimal" knee by providing a relaxation-factor on the response-time distance from the optimal knee, then it is also capable of finding more efficiently a satisfactory knee under these relaxed conditions. We propose two favors of this approach: the first one prunes the VMs using complete plan information received from database engine API, and named as Plan-based Identification of Knee (PIK). On the other hand, to further increase the efficiency of the identification of the knee VM, we propose a sub-plan based pruning algorithm called Sub-Plan-based Identification of Knee (SPIK), which requires modifications in the query optimizer. We have evaluated PIK on a commercial system and found that it often requires processing for only 20% of the total VMs. The efficiency of the algorithm is further increased significantly, by using 10-20% relaxation in response-time. For evaluating SPIK , we prototyped it on an open-source engine { Postgresql 9.3, and also implemented it as Java wrapper program with the commercial engine. Experimentally, the processing done by SPIK is found to be only 40% of the PIK approach. Therefore, from an overall perspective, this thesis facilitates the desired migration of enterprise databases to cloud platforms, by identifying the VM(s) that offer competitive tradeoffs between money and time for the given query.