One of the major research trends currently is the evolution of heterogeneous parallel computing. GP-GPU computing is being widely used and several applications have been designed to exploit the massive parallelism that GP-GPU's have to offer. While GPU's have always been widely used in areas of computer vision for image processing, little has been done to investigate whether the massive parallelism provided by GP-GPU's can be utilized effectively for Natural Language Processing(NLP) tasks. In this work, we investigate and explore the power of GP-GPU's in the task of learning language models. More specifically, we investigate the performance of training Polyglot language models using deep belief neural networks. We evaluate the performance of training the model on the GPU and present optimizations that boost the performance on the GPU.One of the key optimizations, we propose increases the performance of a function involved in calculating and updating the gradient by approximately 50 times on the GPU for sufficiently large batch sizes. We show that with the above optimizations, the GP-GPU's performance on the task increases by factor of approximately 3-4. The optimizations we made are generic Theano optimizations and hence potentially boost the performance of other models which rely on these operations.We also show that these optimizations result in the GPU's performance at this task being now comparable to that on the CPU. We conclude by presenting a thorough evaluation of the applicability of GP-GPU's for this task and highlight the factors limiting the performance of training a Polyglot model on the GPU.
Decolorization is the process to convert a color image or video to its grayscale version, and it has received great attention in recent years. An ideal decolorization algorithm should preserve the original color contrast as much as possible. Meanwhile, it should provide the final decolorized result as fast as possible. However, most of the current methods are suffering from either unsatisfied color information preservation or high computational cost, limiting their application value. In this paper, a simple but effective technique is proposed for real-time decolorization. Based on the typical rgb2gray() color conversion model, which produces a grayscale image by linearly combining R, G, and B channels, we propose a dominant color hypothesis and a corresponding distance measurement metric to evaluate the quality of grayscale conversion. The local optimum scheme provides several "good" candidates in a confidence interval, from which the "best" result can be extracted. Experimental results demonstrate that remarkable simplicity of the proposed method facilitates the process of high resolution images and videos in real-time using a common CPU.
Feature tracking in video is a crucial task in computer vision. Usually, the tracking problem is handled one feature at a time, using a single-feature tracker like the Kanade-Lucas-Tomasi algorithm, or one of its derivatives. While this approach works quite well when dealing with high-quality video and "strong" features, it often falters when faced with dark and noisy video containing low-quality features. We present a framework for jointly tracking a set of features, which enables sharing information between the different features in the scene. We show that our method can be employed to track features for both rigid and nonrigid motions (possibly of few moving bodies) even when some features are occluded. Furthermore, it can be used to significantly improve tracking results in poorly-lit scenes (where there is a mix of good and bad features). Our approach does not require direct modeling of the structure or the motion of the scene, and runs in real time on a single CPU core.
The focus of this paper is speeding up the evaluation of convolutional neural networks. While delivering impressive results across a range of computer vision and machine learning tasks, these networks are computationally demanding, limiting their deployability. Convolutional layers generally consume the bulk of the processing time, and so in this work we present two simple schemes for drastically speeding up these layers. This is achieved by exploiting cross-channel or filter redundancy to construct a low rank basis of filters that are rank-1 in the spatial domain. Our methods are architecture agnostic, and can be easily applied to existing CPU and GPU convolutional frameworks for tuneable speedup performance. We demonstrate this with a real world network designed for scene text character recognition, showing a possible 2.5x speedup with no loss in accuracy, and 4.5x speedup with less than 1% drop in accuracy, still achieving state-of-the-art on standard benchmarks.
We present techniques for speeding up the test-time evaluation of large convolutional networks, designed for object recognition tasks. These models deliver impressive accuracy but each image evaluation requires millions of floating point operations, making their deployment on smartphones and Internet-scale clusters problematic. The computation is dominated by the convolution operations in the lower layers of the model. We exploit the linear structure present within the convolutional filters to derive approximations that significantly reduce the required computation. Using large state-of-the-art models, we demonstrate we demonstrate speedups of convolutional layers on both CPU and GPU by a factor of 2x, while keeping the accuracy within 1% of the original model.
In this note, we show that the L2 optimal solutions to most real multiple view L2 triangulation problems can be efficiently obtained by two-stage Newton-like iterative methods, while the difficulty of such problems mainly lies in how to verify the L2 optimality. Such a working two-stage bundle adjustment approach features: first, the algorithm is initialized by symmedian point triangulation, a multiple-view generalization of the mid-point method; second, a symbolic-numeric method is employed to compute derivatives accurately; third, globalizing strategy such as line search or trust region is smoothly applied to the underlying iteration which assures algorithm robustness in general cases. Numerical comparison with tfml method shows that the local minimizers obtained by the two-stage iterative bundle adjustment approach proposed here are also the L2 optimal solutions to all the calibrated data sets available online by the Oxford visual geometry group. Extensive numerical experiments indicate the bundle adjustment approach solves more than 99% the real triangulation problems optimally. An IEEE 754 double precision C++ implementation shows that it takes only about 0.205 second tocompute allthe 4983 points in the Oxford dinosaur data setvia Gauss-Newton iteration hybrid with a line search strategy on a computer with a 3.4GHz Intel i7 CPU.
This paper presents an approach for employing artificial neural networks (NN) to emulate an ensemble Kalman filter (EnKF) as a method of data assimilation. The assimilation methods are tested in the Simplified Parameterizations PrimitivE-Equation Dynamics (SPEEDY) model, an atmospheric general circulation model (AGCM), using synthetic observational data simulating localization of balloon soundings. For the data assimilation scheme, the supervised NN, the multilayer perceptrons (MLP-NN), is applied. The MLP-NN are able to emulate the analysis from the local ensemble transform Kalman filter (LETKF). After the training process, the method using the MLP-NN is seen as a function of data assimilation. The NN were trained with data from first three months of 1982, 1983, and 1984. A hind-casting experiment for the 1985 data assimilation cycle using MLP-NN were performed with synthetic observations for January 1985. The numerical results demonstrate the effectiveness of the NN technique for atmospheric data assimilation. The results of the NN analyses are very close to the results from the LETKF analyses, the differences of the monthly average of absolute temperature analyses is of order 0.02. The simulations show that the major advantage of using the MLP-NN is better computational performance, since the analyses have similar quality. The CPU-time cycle assimilation with MLP-NN is 90 times faster than cycle assimilation with LETKF for the numerical experiment.
Benchmarking is key for developing and comparing optimization algorithms. In this paper, a CUDA-based real parameter optimization benchmark (cuROB) is introduced. Test functions of diverse properties are included within cuROB and implemented efficiently with CUDA. Speedup of one order of magnitude can be achieved in comparison with CPU-based benchmark of CEC'14.
We present a novel stereo vision algorithm that is capable of obstacle detection on a mobile-CPU processor at 120 frames per second. Our system performs a subset of standard block-matching stereo processing, searching only for obstacles at a single depth. By using an onboard IMU and state-estimator, we can recover the position of obstacles at all other depths, building and updating a full depth-map at framerate. Here, we describe both the algorithm and our implementation on a high-speed, small UAV, flying at over 20 MPH (9 m/s) close to obstacles. The system requires no external sensing or computation and is, to the best of our knowledge, the first high-framerate stereo detection system running onboard a small UAV.
Many diseases cause significant changes to the concentrations of small molecules (aka metabolites) that appear in a person's biofluids, which means such diseases can often be readily detected from a person's "metabolic profile". This information can be extracted from a biofluid's NMR spectrum. Today, this is often done manually by trained human experts, which means this process is relatively slow, expensive and error-prone. This paper presents a tool, Bayesil, that can quickly, accurately and autonomously produce a complex biofluid's (e.g., serum or CSF) metabolic profile from a 1D1H NMR spectrum. This requires first performing several spectral processing steps then matching the resulting spectrum against a reference compound library, which contains the "signatures" of each relevant metabolite. Many of these steps are novel algorithms and our matching step views spectral matching as an inference problem within a probabilistic graphical model that rapidly approximates the most probable metabolic profile. Our extensive studies on a diverse set of complex mixtures, show that Bayesil can autonomously find the concentration of all NMR-detectable metabolites accurately (~90% correct identification and ~10% quantification error), in <5minutes on a single CPU. These results demonstrate that Bayesil is the first fully-automatic publicly-accessible system that provides quantitative NMR spectral profiling effectively -- with an accuracy that meets or exceeds the performance of trained experts. We anticipate this tool will usher in high-throughput metabolomics and enable a wealth of new applications of NMR in clinical settings. Available at http://www.bayesil.ca.
We present a novel, real-time algorithm to track the trajectory of each pedestrian in moderately dense crowded scenes. Our formulation is based on an adaptive particle-filtering scheme that uses a combination of various multi-agent heterogeneous pedestrian simulation models. We automatically compute the optimal parameters for each of these different models based on prior tracked data and use the best model as motion prior for our particle-filter based tracking algorithm. We also use our "mixture of motion models" for adaptive particle selection and accelerate the performance of the online tracking algorithm. The motion model parameter estimation is formulated as an optimization problem, and we use an approach that solves this combinatorial optimization problem in a model independent manner and hence scalable to any multi-agent pedestrian motion model. We evaluate the performance of our approach on different crowd video datasets and highlight the improvement in accuracy over homogeneous motion models and a baseline mean-shift based tracker. In practice, our formulation can compute trajectories of tens of pedestrians on a multi-core desktop CPU in in real time and offer higher accuracy as compared to prior real time pedestrian tracking algorithms.
Functional Magnetic Resonance Imaging (fMRI) is a powerful non-invasive tool for localizing and analyzing brain activity. This study focuses on one very important aspect of the functional properties of human brain, specifically the estimation of the level of parallelism when performing complex cognitive tasks. Using fMRI as the main modality, the human brain activity is investigated through a purely data-driven signal processing and dimensionality analysis approach. Specifically, the fMRI signal is treated as a multi-dimensional data space and its intrinsic `complexity' is studied via dataset fractal analysis and blind-source separation (BSS) methods. One simulated and two real fMRI datasets are used in combination with Independent Component Analysis (ICA) and fractal analysis for estimating the intrinsic (true) dimensionality, in order to provide data-driven experimental evidence on the number of independent brain processes that run in parallel when visual or visuo-motor tasks are performed. Although this number is can not be defined as a strict threshold but rather as a continuous range, when a specific activation level is defined, a corresponding number of parallel processes or the casual equivalent of `cpu cores' can be detected in normal human brain activity.
The considerable mathematical knowledge encoded by the Flyspeck project is combined with external automated theorem provers (ATPs) and machine-learning premise selection methods trained on the proofs, producing an AI system capable of answering a wide range of mathematical queries automatically. The performance of this architecture is evaluated in a bootstrapping scenario emulating the development of Flyspeck from axioms to the last theorem, each time using only the previous theorems and proofs. It is shown that 39% of the 14185 theorems could be proved in a push-button mode (without any high-level advice and user interaction) in 30 seconds of real time on a fourteen-CPU workstation. The necessary work involves: (i) an implementation of sound translations of the HOL Light logic to ATP formalisms: untyped first-order, polymorphic typed first-order, and typed higher-order, (ii) export of the dependency information from HOL Light and ATP proofs for the machine learners, and (iii) choice of suitable representations and methods for learning from previous proofs, and their integration as advisors with HOL Light. This work is described and discussed here, and an initial analysis of the body of proofs that were found fully automatically is provided.
Fetal ECG (FECG) telemonitoring is an important branch in telemedicine. The design of a telemonitoring system via a wireless body-area network with low energy consumption for ambulatory use is highly desirable. As an emerging technique, compressed sensing (CS) shows great promise in compressing/reconstructing data with low energy consumption. However, due to some specific characteristics of raw FECG recordings such as non-sparsity and strong noise contamination, current CS algorithms generally fail in this application. This work proposes to use the block sparse Bayesian learning (BSBL) framework to compress/reconstruct non-sparse raw FECG recordings. Experimental results show that the framework can reconstruct the raw recordings with high quality. Especially, the reconstruction does not destroy the interdependence relation among the multichannel recordings. This ensures that the independent component analysis decomposition of the reconstructed recordings has high fidelity. Furthermore, the framework allows the use of a sparse binary sensing matrix with much fewer nonzero entries to compress recordings. Particularly, each column of the matrix can contain only two nonzero entries. This shows the framework, compared to other algorithms such as current CS algorithms and wavelet algorithms, can greatly reduce code execution in CPU in the data compression stage.
Estimation of Distribution Algorithms (EDAs) require flexible probability models that can be efficiently learned and sampled. Restricted Boltzmann Machines (RBMs) are generative neural networks with these desired properties. We integrate an RBM into an EDA and evaluate the performance of this system in solving combinatorial optimization problems with a single objective. We assess how the number of fitness evaluations and the CPU time scale with problem size and with problem complexity. The results are compared to the Bayesian Optimization Algorithm, a state-of-the-art EDA. Although RBM-EDA requires larger population sizes and a larger number of fitness evaluations, it outperforms BOA in terms of CPU times, in particular if the problem is large or complex. RBM-EDA requires less time for model building than BOA. These results highlight the potential of using generative neural networks for combinatorial optimization.
In large scale machine learning and data mining problems with high feature dimensionality, the Euclidean distance between data points can be uninformative, and Distance Metric Learning (DML) is often desired to learn a proper similarity measure (using side information such as example data pairs being similar or dissimilar). However, high dimensionality and large volume of pairwise constraints in modern big data can lead to prohibitive computational cost for both the original DML formulation in Xing et al. (2002) and later extensions. In this paper, we present a distributed algorithm for DML, and a large-scale implementation on a parameter server architecture. Our approach builds on a parallelizable reformulation of Xing et al. (2002), and an asynchronous stochastic gradient descent optimization procedure. To our knowledge, this is the first distributed solution to DML, and we show that, on a system with 256 CPU cores, our program is able to complete a DML task on a dataset with 1 million data points, 22-thousand features, and 200 million labeled data pairs, in 15 hours; and the learned metric shows great effectiveness in properly measuring distances.
The primal-dual column generation method (PDCGM) is a general-purpose column generation technique that relies on the primal-dual interior point method to solve the restricted master problems. The use of this interior point method variant allows to obtain suboptimal and well-centered dual solutions which naturally stabilizes the column generation. As recently presented in the literature, reductions in the number of calls to the oracle and in the CPU times are typically observed when compared to the standard column generation, which relies on extreme optimal dual solutions. However, these results are based on relatively small problems obtained from linear relaxations of combinatorial applications. In this paper, we investigate the behaviour of the PDCGM in a broader context, namely when solving large-scale convex optimization problems. We have selected applications that arise in important real-life contexts such as data analysis (multiple kernel learning problem), decision-making under uncertainty (two-stage stochastic programming problems) and telecommunication and transportation networks (multicommodity network flow problem). In the numerical experiments, we use publicly available benchmark instances to compare the performance of the PDCGM against recent results for different methods presented in the literature, which were the best available results to date. The analysis of these results suggests that the PDCGM offers an attractive alternative over specialized methods since it remains competitive in terms of number of iterations and CPU times even for large-scale optimization problems.
We discuss several strategies to implement Dykstra's projection algorithm on NVIDIA's compute unified device architecture (CUDA). Dykstra's algorithm is the central step in and the computationally most expensive part of statistical multi-resolution methods. It projects a given vector onto the intersection of convex sets. Compared with a CPU implementation our CUDA implementation is one order of magnitude faster. For a further speed up and to reduce memory consumption we have developed a new variant, which we call incomplete Dykstra's algorithm. Implemented in CUDA it is one order of magnitude faster than the CUDA implementation of the standard Dykstra algorithm. As sample application we discuss using the incomplete Dykstra's algorithm as preprocessor for the recently developed super-resolution optical fluctuation imaging (SOFI) method (Dertinger et al. 2009). We show that statistical multi-resolution estimation can enhance the resolution improvement of the plain SOFI algorithm just as the Fourier-reweighting of SOFI. The results are compared in terms of their power spectrum and their Fourier ring correlation (Saxton and Baumeister 1982). The Fourier ring correlation indicates that the resolution for typical second order SOFI images can be improved by about 30 per cent. Our results show that a careful parallelization of Dykstra's algorithm enables its use in large-scale statistical multi-resolution analyses.
Deep convolutional neural networks (CNN) have recently been shown in many computer vision and pattern recog- nition applications to outperform by a significant margin state- of-the-art solutions that use traditional hand-crafted features. However, this impressive performance is yet to be fully exploited in robotics. In this paper, we focus one specific problem that can benefit from the recent development of the CNN technology, i.e., we focus on using a pre-trained CNN model as a method of generating an image representation appropriate for visual loop closure detection in SLAM (simultaneous localization and mapping). We perform a comprehensive evaluation of the outputs at the intermediate layers of a CNN as image descriptors, in comparison with state-of-the-art image descriptors, in terms of their ability to match images for detecting loop closures. The main conclusions of our study include: (a) CNN-based image representations perform comparably to state-of-the-art hand- crafted competitors in environments without significant lighting change, (b) they outperform state-of-the-art competitors when lighting changes significantly, and (c) they are also significantly faster to extract than the state-of-the-art hand-crafted features even on a conventional CPU and are two orders of magnitude faster on an entry-level GPU.
We propose a simple two-step approach for speeding up convolution layers within large convolutional neural networks based on tensor decomposition and discriminative fine-tuning. Given a layer, we use non-linear least squares to compute a low-rank CP-decomposition of the 4D convolution kernel tensor into a sum of a small number of rank-one tensors. At the second step, this decomposition is used to replace the original convolutional layer with a sequence of four convolutional layers with small kernels. After such replacement, the entire network is fine-tuned on the training data using standard backpropagation process. We evaluate this approach on two CNNs and show that it is competitive with previous approaches, leading to higher obtained CPU speedups at the cost of lower accuracy drops for the smaller of the two networks. Thus, for the 36-class character classification CNN, our approach obtains a 8.5x CPU speedup of the whole network with only minor accuracy drop (1% from 91% to 90%). For the standard ImageNet architecture (AlexNet), the approach speeds up the second convolution layer by a factor of 4x at the cost of $1\%$ increase of the overall top-5 classification error.