China
Africa
Explore chapters and articles related to this topic
Machine Learning
Published in Seyedeh Leili Mirtaheri, Reza Shahbazian, Machine Learning Theory to Applications, 2022
Seyedeh Leili Mirtaheri, Reza Shahbazian
Various learning algorithms such as principal components analysis and cluster analysis seek better representations of input data through the training step. Feature learning algorithms, or representation learning algorithms, generally try to keep the input information and also modify it to make it useful. It is usually done as a pre-processing step before classification or predictions. This modification allows the algorithm to reconstruct a set of input data that comes from the unknown data-generating distributions without staying necessarily faithful to settings that are unacceptable under that distribution. After that, there is no need for manual feature engineering, but the machine can also learn the features and perform the desired task. The feature learning system employs a set of techniques to automatically find the hidden pattern of features in raw data and classify them.
Deep Learning
Published in Peter Wlodarczak, Machine Learning and its Applications, 2019
Feature learning is a set of techniques that allows a machine to automatically detect features without the need of a feature engineer to manually extract the features of interest. For instance, traditional approaches for object recognition in an image require a feature engineer to isolate the area in the image where the object is located and to omit the background. Using deep architectures, the object of interest is automatically detected, a process called object detection. A deep architecture is capable of doing both, object detection and object recognition. The layered architecture of a deep learner creates higher abstractions of the input in every layer. The output of every layer is a lower dimensional projection of the input. Provided the deep network is optimally weighted, the output of the network is a high-level abstraction of the raw data representing an automatic feature set. This feature set can then be used as input for the actual classification task. Automatic feature detection also helps to avoid human bias. Using, for instance, a convolutional neural network, the convolutional layer acts as a feature extractor.
Laser scanning for bridge inspection
Published in Belén Riveiro, Roderik Lindenbergh, Laser Scanning, 2019
Linh Truong-Hong, Debra F. Laefer
Demonstration of extracting surface defects due to chemical attack, corrosion, and water bleed through texture information of a point cloud proves machine learning–based methods can detect such defects with high accuracy. However, attributes of the point cloud including backscatter energy (or intensity value) and RGB colors depend on various factors, for example, a laser scanning unit (e.g. a wavelength or divergence of a laser beam) and a camera, properties of an object’s surface (e.g. material, color, and roughness of the surface), scanning range and incident angle, and atmospheric conditions (e.g. humidity, temperature, and wind). As such, values of intensity and RGB colors of damaged surfaces vary in a wide range, since the structure’s surface is to be scanned in different periods of time. Clearly, these values also change in different structures. As such, the machine learning methods can give reliable results once a training data set is large enough and sufficient features are to be prepared. That implies a numerous surface damage to be captured, and feature engineering may require huge labor time to label the point cloud as either “damaged” or “nondamaged”. A robust method should be developed to obtain invariant values of the features from various sources of laser scanning data. As only two main features are intensity and RGB colors of the point cloud derived directly from the TLS, these features may not be enough for constructing a good generalization model. Additional features used in the training model can improve the performance in term of classification accuracy of the predicted model. Automated feature learning can be achieved by using a deep learning method, but this learning method requires huge training data.
Stacking ensemble transfer learning based thermal displacement prediction system
Published in International Journal of Optomechatronics, 2023
Ping-Huan Kuo, Chia-Ho Lee, Her-Terng Yau
Deep learning is a small set in the field of machine learning, where the difference between these two learning structures is that deep learning will learn how to extract data features by itself but machine learning not. The artificial neural network algorithm are commonly used in feature learning, and there have been several frameworks based on artificial neural network, such as convolutional neural network, deep belief network, etc. In addition, there is certain progress in the modern semiconductor industry, Meada et al. have verified through tests that the time is significantly speed up while GPU is applied on the calculation of neural networks.[8] Therefore, deep learning has been widely used in academia and industries in recent years, and many problems either difficult or time consuming have been solved one by one. As a result, in this study, it is intended to predict the thermal error of spindle by deep learning, thereby pushing the numerical control machine tools moving forward to the era of intelligence.
Blood transfusion prediction using restricted Boltzmann machines
Published in Computer Methods in Biomechanics and Biomedical Engineering, 2020
Jenny Cifuentes, Yuanyuan Yao, Min Yan, Bin Zheng
Ongoing work in this area includes an in-deep analysis of Deep Learning architectures implementation, which have achieved successful results in different biomedical applications. In this context, autoencoders and Convolutional Neural Networks can be explored for the feature learning process. Likewise, in case of datasets that include time sequences to characterize clinical variables, Long Short Term Memory (LSTM) Neural Networks could be explored. These approaches have a strong potentiality in the blood transfusion prediction area. Likewise, there is a growing interest among the medical community in interpreting machine learning models and gaining insights into their inner working calculations to mainly support the decision-making process. In this context, the development of strategies that allow to understand this process for machine learning algorithms, which have reported high classification rates, is paramount.
Machine learning in drying
Published in Drying Technology, 2020
Also known as feature learning, representation learning relies on methods allowing a computer to automatically discover the representations needed for detection or classification from raw data.[38] The widely acclaimed deep learning models based on artificial neural networks (ANN) is the best example of representation learning. Representation learning algorithms could be supervised or unsupervised. Considering that ANNs can be used for modeling, they have been leveraged as a basic tool in the process industry for process monitoring, fault classification, and soft sensor modeling.[5,39]