China
Africa
Explore chapters and articles related to this topic
Covariate Shift in Machine Learning
Published in Monika Mangla, Subhash K. Shinde, Vaishali Mehta, Nonita Sharma, Sachi Nandan Mohanty, Handbook of Research on Machine Learning, 2022
Santosh Chapaneri, Deepak Jayaswal
The FW optimization concept has been applied in the literature for various applications: (i) matrix factorization where the convergence was shown using the approximation quality obtained with the FW duality gap [9], (ii) image and video co-localization formulated using FW leading to an improved computational efficiency [10], (iii) structured support vector machine (SSVM) with the SVM duality gap equivalent to the FW duality gap [11], etc. To remove the influence of “bad” visited vertices, the away-steps FW algorithm [18] was shown to converge at a linear convergence rate in [19].
Aggregating multi-scale contextual features from multiple stages for semantic image segmentation
Published in Connection Science, 2021
Dingchao Jiang, Hua Qu, Jihong Zhao, Jianlong Zhao, Meng-Yen Hsieh
Probabilistic graphical models, such as conditional random fields, are used in some models. CRF is a probability model that describes the dependencies between nodes (Lafferty et al., 2001) to capture spatial information for classification. Krähenbühl and Koltun (2011) introduced a fully connected CRF model, in which the pairwise edge potentials are defined by a linear combination of Gaussian kernels on a complete set of pixels. Liu et al. (2015) used a CNN that is trained on the ImageNet dataset to construct potentials of superpixels and learns CRF parameters by using a structured support vector machine. CRFasRNN (Zheng et al., 2015) reformulates dense CRFs as recurrent neural networks and trains the whole network end-to-end with a back-propagation algorithm. Arnab et al. (2016) employed higher-order potentials in a CRF to improve segmentation performance. DeepLab frameworks (Chen, Papandreou, Kokkinos, et al., 2017) use CRFs for post-processing. Colovic et al. (2017) proposed a CNN+CRFs framework to learn arbitrary pairwise CRFs potentials. The fully convolutional residual continuous CRF network (L. Zhou et al., 2020) has three subnetworks, namely, a unary network, a pairwise network, and a superpixel-based continuous conditional random field network to fuse multi-scale features and generate full spatial resolution predictions.