Number Of Neural Network Architectures Research Articles

Influence of Neural Network Receptive Field on Monocular Depth and Ego-Motion Estimation

We present an analysis of a self-supervised learning approach for monocular depth and ego-motion estimation. This is an important problem for computer vision systems of robots, autonomous vehicles and other intelligent agents, equipped only with monocular camera sensor. We have explored a number of neural network architectures that perform single-frame depth and multi-frame camera pose predictions to minimize photometric error between consecutive frames on a sequence of camera images. Unlike other existing works, our proposed approach called ERF-SfMLearner examines the influence of the deep neural network receptive field on the performance of depth and ego-motion estimation. To do this, we study the modification of network layers with two convolution operators with extended receptive field: dilated and deformable convolutions. We demonstrate on the KITTI dataset that increasing the receptive field leads to better metrics and lower errors both in terms of depth and ego-motion estimation. Code is publicly available at github.com/linukc/ERF-SfMLearner.

SEA-net: Sequence attention network for seismic event detection and phase arrival picking

Automatic seismic phase detection and phase arrival time picking are always fundamental tasks in seismology. With the emergence of machine learning, and deep learning in particular, a number of neural network architectures are being used to explore the correlations between seismic waveforms and hidden information of interest. In this paper, we present a sequential attention (SEA) module that can learn both local and long-distance corrections to effectively extract sequential features. This module is lightweight and plug-in-play. Based on the SEA module, a fully convolutional neural network with sequential attention mechanism was constructed to detect seismic events and determine the arrival time of phases. Further experiments on STEAD dataset, in comparison with existing methods, show that our proposed model outperforms other algorithms with a higher F1-score (100% for seismic detection, 98.8% for P arrival picking, and 89.0% for S arrival picking) and fewer residual errors. Comparatively, our model has a quarter of the model parameters and computational complexity as well as a 36.7% reduction in processing time compared to the current state-of-the-art method. This demonstrates that the proposed architecture is effective for seismic sequential learning tasks and, as a result, can serve as an improved alternative backbone.

Research of U-Net-Based CNN Architectures for Metal Surface Defect Detection

The quality, wear and safety of metal structures can be controlled effectively, provided that surface defects, which occur on metal structures, are detected at the right time. Over the past 10 years, researchers have proposed a number of neural network architectures that have shown high efficiency in various areas, including image classification, segmentation and recognition. However, choosing the best architecture for this particular task is often problematic. In order to compare various techniques for detecting defects such as “scratch abrasion”, we created and investigated U-Net-like architectures with encoders such as ResNet, SEResNet, SEResNeXt, DenseNet, InceptionV3, Inception-ResNetV2, MobileNet and EfficientNet. The relationship between training validation metrics and final segmentation test metrics was investigated. The correlation between the loss function, the DSC, IoU, Recall, Precision and F1 validation metrics and DSC test metrics was calculated. Recognition accuracy was analyzed as affected by the optimizer during neural network training. In the context of this problem, neural networks trained using the stochastic gradient descent optimizer with Nesterov momentum were found to have the best generalizing properties. To select the best model during its training on the basis of the validation metrics, the main test metrics of recognition quality (Dice similarity coefficient) were analyzed depending on the validation metrics. The ResNet and DenseNet models were found to achieve the best generalizing properties for our task. The highest recognition accuracy was attained using the U-Net model with a ResNet152 backbone. The results obtained on the test dataset were DSC=0.9304 and IoU=0.9122.

Selection of the optimal neural network of a Smart building system for object recognition

Recently, artificial intelligence has captured a number of solutions that help us in everyday life. It will not surprise anyone now. It is in every modern smartphone and helps us to search, sort and process information faster. Today, smart home systems have taken a significant place among the developments in the field of information technology. The huge amount of data generated in them and the variety of formats of this data does not allow to create a universal mechanism for their productive processing. Therefore, the integration of neural networks, which is best suited for individual tasks, will provide us with high efficiency of smart home systems with minimal errors in decisions. Year by year, the interest in solving more complex tasks of object recognition is growing, due to automation needs for shaped communication processes in intelligent systems. Therefore, improving the implementation of the recognition of computer image systems is relevant. One of the promising directions for solving this problem is based on the use of artificial neural networks and neurocomputers as the most progressive in relation to the problems of classification of pattern recognition tasks. In our time, a large number of neural network architectures are proposed for application in the recognition of objects. The analysis of the proposed solutions shows that there is still no such model that would be the best among all the resulting performance parameters. Prospects for the improvement of architecture are seen in convolutional neural networks. The advantages of roller networks over multilayers are to use a common weight in the roller coasters, which means that for each pixel of the layer is used the same filter.

Entity-Level Classification of Adverse Drug Reaction: A Comparative Analysis of Neural Network Models

An experimental work on the analysis of effectiveness of neural network models applied to the classification of adverse drug reactions at the entity level is described. Aspect-level sentiment analysis, which aims to determine the sentimental class of a specific aspect conveyed in user opinions, has been actively studied for more than 10 years. A number of neural network architectures have been proposed. Even though the models based on these architectures have much in common, they differ in certain components. In this paper, the applicability of the neural network models developed for the aspect-level sentiment analysis to the problem of the classification of adverse drug reactions is studied. Extensive experiments on English language texts of biomedical topic, including health records, scientific literature, and social media have been conducted. The proposed models mentioned above are compared with one of the best model based on the support vector machine method and a large set of features.

Machine tool positioning error compensation using artificial neural networks

This paper is a study of the application of artificial neural networks to the problem of calculating error compensation values for axis positioning on a machine tool. The primary focus is on the development of a neural network-based system that could be implemented and integrated into the open architecture control system of an actual machine. A number of neural network architectures were examined for applicability to the problem and one was selected and implemented on the actual machine. Positioning error compensation capabilities were evaluated using industry standard equipment and procedures, and the results obtained were compared with the capabilities of standard error compensation routines in machine tool controls.

Soft computing models based prediction of cutting speed and surface roughness in wire electro-discharge machining of tungsten carbide cobalt composite

In the present study, a second order multi-variable regression model and a feed-forward back-propagation neural network (BPNN) model have been developed to correlate the input process parameters, such as pulse on-time, pulse off-time, peak current, and capacitance with the performance measures namely, cutting speed and surface roughness while wire electro-discharge machining (WEDM) of tungsten carbide-cobalt (WC-Co) composite material. From a large number of neural network architectures, 4-11-2 has been found to be the optimal one, which can predict cutting speed and surface roughness with 3.29% overall mean prediction error. The multivariable regression model yields an overall mean prediction error of 6.02%. Both the models have been used to study the effect of input parameters on the cutting speed and surface roughness, and finally to corroborate them with those of the experimental results. Scanning electron micrographs reveal that at higher energy level the machined surface is characterized by several microcracks and loosely bound solidified WC grains.

The comparison of different feed forward neural network architectures for ECG signal diagnosis

The electrocardiograms (ECGs) record the electrical activity of the heart and are used to diagnose many heart disorders. This paper proposes a two-stage feed forward neural network for ECG signal classification. The research is aimed at the design of an intelligent ECG diagnosis tool that can recognise heart abnormalities while reducing the complexity, cost, and response time of the system. A number of neural network architectures are designed and compared for their ability to classify six different heart conditions. Two network architectures based on one stage and two stage feed forward neural networks are chosen for this investigation. The training and testing ECG signals are obtained from MIT-BIH database. The network inputs are comprised of 12 ECG features and 13 compressed components of each heart beat signal. The performance of the different modules as well as the efficiency of the whole system is presented. Among different architectures, a proposed multi-stage network named NET_BST possesses the highest recognition rate of around 93%. Therefore, this network proves to be a suitable candidate in ECG signal diagnosis systems.