China
Africa
Explore chapters and articles related to this topic
Applications of AI in Medical Science and Drug Development
Published in Mark Chang, Artificial Intelligence for Drug Development, Precision Medicine, and Healthcare, 2020
Deep learning networks such as CNNs (Avendi et al., 2016) and DBNs (Carneiro et al., 2012; Ngo et al., 2016) have recently been used in cardiac imaging, including left/right ventricle segmentation (Zhen et al., 2016), retinal vessel segmentation (Wang et al., 2015; Chandrakumar and Kathirvel, 2016). A new convolutional deep belief network is proposed by Zhen et al. (2016) for direct estimation of a ventricular volume from images without performing segmentation at all. Other work in deep learning for CV can be found: a system that automatically learns the most effective application-specific vesselness measurement from an expert-annotated dataset (Zheng et al., 2011, 2013), a CNN for pixel-wise classification of retinal vessels (Wang et al., 2015; Wu et al., 2016), deep learning for detecting retinal vessel microaneurysms (Haloi, 2015), and the classification of diabetic retinopathy (Chandrakumar and Kathirvel, 2016) via fundus imaging.
Deep Learning: Fundamentals and Beyond
Published in Mayank Vatsa, Richa Singh, Angshul Majumdar, Deep Learning in Biometrics, 2018
Shruti Nagpal, Maneet Singh, Mayank Vatsa, Richa Singh
Over the past several years, progress has also been made with RBMs in the form of incorporating unsupervised regularizers and modifying them for different applications. In 2008, Lee et al. presented sparse DBNs [33], mimicking certain properties of the human brain’s visual area, V2. A regularization term is added to the loss function of an RBM to introduce sparsity in the learned representations. Similar to the performance observed with stacked autoencoders, the first layer was seen to learn edge filters (like the Gabor filters), and the second layer encoded correlations of the first-layer responses in the data, along with learning corners and junctions. Following this, a convolutional deep belief network (CDBN) was proposed by Lee et al. [34] for addressing several visual-recognition tasks. The model incorporated a novel probabilistic max-pooling technique for learning hierarchical features from unlabeled data. CDBN is built using the proposed convolutional RBMs, which incorporate convolution in the feature learning process of traditional RBMs. Probabilistic max-pooling is used at the time of stacking convolutional RBMs to create CDBNs for learning hierarchical representations. To eliminate trivial solutions, sparsity has also been enforced on the hidden representations. Inspired by the observation that both coarse and fine details of images may provide discriminative information for image classification, Tang and Mohamad proposed multiresolution DBNs [35]. The model used multiple independent RBMs trained on different levels of the Laplacian pyramid of an image and combined the learned representations to create the input to a final RBM. This entire model is known as multiresolution DBN, and the objective is to extract meaningful representations from different resolutions of the given input image. Coarse and fine details of the input are used for feature extraction, thereby enabling the proposed model to encode multiple variations. Further, in 2014, in an attempt to model the intermodality variations for a multimodal classification task, Srivastava and Salakhutdinov proposed the multimodal DBM [36]. The model aimed to learn a common (joint) representation for samples belonging to two different modalities such that the learned feature is representative of both the samples. The model also ensures that it is able to generate a common representation given only a sample from a single modality. In the proposed model, two DBMs are trained for two modalities, followed by a DBM trained on the combined learned representations from the two previous DBMs. The learned representation from the third DBM corresponds to the joint representation of the two modalities. Recently, Huang et al. proposed an RBM-based model for unconstrained multimodal multilablel learning [37] termed a multilabel conditional RBM. It aims to learn a joint feature representation over multiple modalities and predict multiple labels.
Deep Learning: Differential Privacy Preservation in the Era of Big Data
Published in Journal of Computer Information Systems, 2023
Recent advances in the internet of things (IoT) are developed with tremendous technologies in communication. The large in IoT is also referred to as big data; hence to facilitate feature learning in big data, Qingchen Zhang et al.49 proposed a double-projection DCM (DPDCM). That proposed model adopts double projection layers instead of hidden layers in the DCM. A learning algorithm was used for the training of the projection layer. Moreover, cloud computing was adopted to enhance training accuracy through crowdsourcing data on the cloud. A privacy-preserving model was suggested based on BGV encryption to improve data privacy. Due to the remarkable developments of DL in the medicine and healthcare domain, NhatHai Phan et al.50 proposed a private convolutional deep belief network (pCDBN). The DBN was designed by -differential privacy and Chebyshev expansion in that approach.
Incorporating travel time means and standard deviations into transportation network design problem: a hybrid method based on column generation and Lagrangian relaxation
Published in Transportation Letters, 2023
Many heuristic approaches such as Bee colony optimization (Nikolić and Teodorović 2013, 2014), Bee colony optimization (Veluscek et al. 2015), genetic algorithm (Xu, Wei, and Hu 2009), variable fixing heuristics (Guimaraes, de Sousa, and Prata 2022), and trial and error algorithm (Lin 2022) were proposed for the NDPs. For global optimization problems, Deng et al. (2022b) developed an enhanced evolutionary algorithm with multiple strategies. Zhao et al. (2022) proposed a novel vibration amplitude spectrum imaging feature extraction method and a new convolutional deep belief network. Song et al. (2023) developed an adaptive cooperative co-evolutionary differential evolution to solve the train scheduling problem. An adaptive differential evolution algorithm was proposed by Deng et al. (2022a). For the airport taxiway planning problem, Deng et al. (2022c) proposed a hybrid method based on particle swarm and ant colony algorithms. Additionally, some studies focused on linear programs (Murty 1980) by employing exact methods. Hulagu and Celikoglu (2022b) studied the electric vehicle routing problem with intermediate nodes. Furthermore, Hulagu and Celikoglu (2022a) located charging stations for electric vehicle routing problems.
COVID-19 diagnosis prediction using classical-to-quantum ensemble model with transfer learning for CT scan images
Published in The Imaging Science Journal, 2021
Wenqian Li, Xing Deng, Haorong Zhao, Haijian Shao, Yingtao Jiang
Moreover, the study shows that the ensemble learning method is better than an individual model in terms of prediction and can effectively prevent over-fitting [37]. Frazao et al. [38] proposed a weighted CNN ensemble method, which weighted the output probability of each network. It is found to be better than the unweighted average method. To rapidly detect the novel coronavirus COVID-19, Zhou et al. [39] used the pre-trained model AlexNet, GoogleNet and ResNet to extract chest CT images features, and finally obtained the ensemble classifier that an ensemble deep learning model (EDL_COVID) via relative majority voting. To classify COVID-19 from chest CT scan and Lung X-ray, Berliana et al. [40] presented an ensemble learning model that built with two levels of learning, namely the base-learners (use Support Vector Classification (SVC), Random Forest (RF) and K-Nearest Neighbours (KNN)) and meta-learner that use SVC. Kundu et al. [41] integrated the three models of Inception v3, ResNet34 and DenseNet201, and used a bagging ensemble technique to detect COVID-19 through chest CT-scan images. Mohammed et al. [42] employed an integrated method for selecting the optimal deep learning based on a novel crow swarm optimization for COVID-19 diagnosis. This method utilized 746 CT images, 632 positive cases of COVID-19 that unimproved CT images of the lung. Ibrahim et al. [43] developed an effective hybrid deep learning model for COVID-19 patterns identification using CT images and obtained a 95% accuracy rate, which was combined of three pre-trained models, namely VGGNet, convolutional deep belief network (CDBN) and high-resolution network (HRNet).