Hybrid pooling with wavelets for convolutional neural networks

The need to detect and classify objects correctly is a constant challenge, being able to recognize them at different scales and scenarios, sometimes cropped or badly lit is not an easy task. Convolutional neural networks (CNN) have become a widely applied technique since they are completely trainable and suitable to extract features. However, the growing number of convolutional neural networks applications constantly pushes their accuracy improvement. Initially, those improvements involved the use of large datasets, augmentation techniques, and complex algorithms. These methods may have a high computational cost. Nevertheless, feature extraction is known to be the heart of the problem. As a result, other approaches combine different technologies to extract better features to improve the accuracy without the need of more powerful hardware resources. In this paper, we propose a hybrid pooling method that incorporates multiresolution analysis within the CNN layers to reduce the feature map size without losing details. To prevent relevant information from losing during the downsampling process an existing pooling method is combined with wavelet transform technique, keeping those details "alive" and enriching other stages of the CNN. Achieving better quality characteristics improves CNN accuracy. To validate this study, ten pooling methods, including the proposed model, are tested using four benchmark datasets. The results are compared with four of the evaluated methods, which are also considered as the state-of-the-art.

Tiled Sparse Coding in Eigenspaces for Image Classification

The automation in the diagnosis of medical images is currently a challenging task. The use of Computer Aided Diagnosis (CAD) systems can be a powerful tool for clinicians, especially in situations when hospitals are overflowed. These tools are usually based on artificial intelligence (AI), a field that has been recently revolutionized by deep learning approaches. blackThese alternatives usually obtain a large performance based on complex solutions, leading to a high computational cost and the need of having large databases. In this work, we propose a classification framework based on sparse coding. Images are blackfirst partitioned into different tiles, and a dictionary is built after applying PCA to these tiles. The original signals are then transformed as a linear combination of the elements of the dictionary. blackThen, they are reconstructed by iteratively deactivating the elements associated with each component. Classification is finally performed employing as features the subsequent reconstruction errors. Performance is evaluated in a real context where distinguishing between four different pathologies: control versus bacterial pneumonia versus viral pneumonia versus COVID-19. blackOur system differentiates between pneumonia patients and controls with an accuracy of 97.74%, whereas in the 4-class context the accuracy is 86.73%. The excellent results and the pioneering use of sparse coding in this scenario evidence that our proposal can assist clinicians when their workload is high.

Transferability of Deep Learning Models for Focus Quality Assessment in Digital Pathology

Out-of-focus sections of whole slide images are a significant source of false positives and other systematic errors in clinical diagnoses. As a result, focus quality assessment (FQA) methods must be able to quickly and accurately differentiate between focus levels in a scan. Recently, deep learning methods using convolutional neural networks (CNNs) have been adopted for FQA. However, the biggest obstacles impeding their wide usage in clinical workflows are their generalizability across different test conditions and their potentially high computational cost. In this study, we focus on the transferability and scalability of CNN-based FQA approaches. We carry out an investigation on ten architecturally diverse networks using five datasets with stain and tissue diversity. We evaluate the computational complexity of each network and scale this to realistic applications involving hundreds of whole slide images. We assess how well each full model transfers to a separate, unseen dataset without fine-tuning. We show that shallower networks transfer well when used on small input patch sizes, while deeper networks work more effectively on larger inputs. Furthermore, we introduce neural architecture search (NAS) to the field and learn an automatically designed low-complexity CNN architecture using differentiable architecture search which achieved competitive performance relative to established CNNs.

Design of 2D Planar Sparse Binned Arrays Based on the Coarray Analysis

The analysis of the beampattern is the base of sparse arrays design process. However, in the case of bidimensional arrays, this analysis has a high computational cost, turning the design process into a long and complex task. If the imaging system development is considered a holistic process, the aperture is a sampling grid that must be considered in the spatial domain through the coarray structure. Here, we propose to guide the aperture design process using statistical parameters of the distribution of the weights in the coarray. We have studied three designs of sparse matrix binned arrays with different sparseness degrees. Our results prove that there is a relationship between these parameters and the beampattern, which is valuable and improves the array design process. The proposed methodology reduces the computational cost up to 58 times with respect to the conventional fitness function based on the beampattern analysis.

Iterative machine learning applied to annotation of text datasets

The purpose of this paper is to analyze the use of different machine learning approaches and algorithms to be integrated as an automated assistance on a tool to aid the creation of new annotated datasets. We evaluate how they scale in an environment without dedicated machine learning hardware. In particular, we study the impact over a dataset with few examples and one that is being constructed. We experiment using deep learning algorithms (Bert) and classical learning algorithms with a lower computational cost (W2V and Glove combined with RF and SVM). Our experiments show that deep learning algorithms have a performance advantage over classical techniques. However, deep learning algorithms have a high computational cost, making them inadequate to an environment with reduced hardware resources. Simulations using Active and Iterative machine learning techniques to assist the creation of new datasets are conducted. For these simulations, we use the classical learning algorithms because of their computational cost. The knowledge gathered with our experimental evaluation aims to support the creation of a tool for building new text datasets.

A Bayesian multiscale CNN framework to predict local stress fields in structures with microscale features

Multiscale computational modelling is challenging due to the high computational cost of direct numerical simulation by finite elements. To address this issue, concurrent multiscale methods use the solution of cheaper macroscale surrogates as boundary conditions to microscale sliding windows. The microscale problems remain a numerically challenging operation both in terms of implementation and cost. In this work we propose to replace the local microscale solution by an Encoder-Decoder Convolutional Neural Network that will generate fine-scale stress corrections to coarse predictions around unresolved microscale features, without prior parametrisation of local microscale problems. We deploy a Bayesian approach providing credible intervals to evaluate the uncertainty of the predictions, which is then used to investigate the merits of a selective learning framework. We will demonstrate the capability of the approach to predict equivalent stress fields in porous structures using linearised and finite strain elasticity theories.

A Nonlinear UAV Control Tuning Under Communication Delay using HPC Strategies in Parameters Space

In practical applications, the presence of delays can deteriorate the performance of the control system or even cause plant instability. However, by properly controlling these delays, it is possible to improve the performance of the mechanism. The present work is based on a proposal to analyze the asymptotic stability and convergence of a quadrotor robot, an unmanned aerial vehicle (UAV), on the performance of a given task, under time delay in the data flow. The effects of the communication delay problem, as well as the response-signal behavior of the quadrotors in the accomplishment of positioning mission are presented and analyzed from the insertion of fixed time delay intervals in the UAVs' data collected by its sensors system. Due to the large search space in the set of parameter combinations and the high computational cost required to perform such an analysis by sequentially executing thousands of simulations, this work proposes an open source GPU-based implementation to simulate the robot behavior. Experimental results show a speedup up to 4900x in comparison to MATLAB® implementation. The implement is available in Colab Google platform.

Nearest Neighbors Search Using Multi-GPU

Meshless methods are increasingly gaining space in the study of electromagnetic phenomena as an alternative to traditional mesh-based methods. One of their biggest advantages is the absence of a mesh to describe the simulation domain. Instead, the domain discretization is done by spreading nodes along the domain and its boundaries. Thus, meshless methods are based on the interactions of each node with all its neighbors, and determining the neighborhood of the nodes becomes a fundamental task. The k-nearest neighbors (kNN) is a well-known algorithm used for this purpose, but it becomes a bottleneck for these methods due to its high computational cost. One of the alternatives to reduce the kNN high computational cost is to use spatial partitioning data structures (e.g., planar grid) that allow pruning when performing the k-nearest neighbors search. Furthermore, many of these strategies employed for kNN search have been adapted for graphics processing units (GPUs) and can take advantage of its high potential for parallelism. Thus, this paper proposes a multi-GPU version of the grid method for solving the kNN problem. It was possible to achieve a speedup of up to 1.99x and up to 3.94x using two and four GPUs, respectively, when compared against the single-GPU version of the grid method.

Deriving Large-Scale Coastal Bathymetry from Sentinel-2 Images Using an HIGH-Performance Cluster: A Case Study Covering North Africa’s Coastal Zone

Coasts are areas of vitality because they host numerous activities worldwide. Despite their major importance, the knowledge of the main characteristics of the majority of coastal areas (e.g., coastal bathymetry) is still very limited. This is mainly due to the scarcity and lack of accurate measurements or observations, and the sparsity of coastal waters. Moreover, the high cost of performing observations with conventional methods does not allow expansion of the monitoring chain in different coastal areas. In this study, we suggest that the advent of remote sensing data (e.g., Sentinel 2A/B) and high performance computing could open a new perspective to overcome the lack of coastal observations. Indeed, previous research has shown that it is possible to derive large-scale coastal bathymetry from S-2 images. The large S-2 coverage, however, leads to a high computational cost when post-processing the images. Thus, we develop a methodology implemented on a High-Performance cluster (HPC) to derive the bathymetry from S-2 over the globe. In this paper, we describe the conceptualization and implementation of this methodology. Moreover, we will give a general overview of the generated bathymetry map for NA compared with the reference GEBCO global bathymetric product. Finally, we will highlight some hotspots by looking closely to their outputs.

FPL: fast Presburger arithmetic through transprecision

Presburger arithmetic provides the mathematical core for the polyhedral compilation techniques that drive analytical cache models, loop optimization for ML and HPC, formal verification, and even hardware design. Polyhedral compilation is widely regarded as being slow due to the potentially high computational cost of the underlying Presburger libraries. Researchers typically use these libraries as powerful black-box tools, but the perceived internal complexity of these libraries, caused by the use of C as the implementation language and a focus on end-user-facing documentation, holds back broader performance-optimization efforts. With FPL, we introduce a new library for Presburger arithmetic built from the ground up in modern C++. We carefully document its internal algorithmic foundations, use lightweight C++ data structures to minimize memory management costs, and deploy transprecision computing across the entire library to effectively exploit machine integers and vector instructions. On a newly-developed comprehensive benchmark suite for Presburger arithmetic, we show a 5.4x speedup in total runtime over the state-of-the-art library isl in its default configuration and 3.6x over a variant of isl optimized with element-wise transprecision computing. We expect that the availability of a well-documented and fast Presburger library will accelerate the adoption of polyhedral compilation techniques in production compilers.