An Improved Hashing Method for Image Retrieval Based on Deep Neural Networks

Qiu, Chengfeng; Wang, Weidong; Lee, Feifei

doi:10.1145/3297067.3297092

Cited by 61 publications

(84 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…∀ : for all | ∈: in | s. t.: such that in many physics-enforced neural network studies (Wang, Teng, & Perdikaris, 2021;Wight & Zhao, 2020), across multitask learning problems (Caruana, 1997;Chen et al, 2018), and evidenced by the emergence of strategies that target learning under competing objectives (Elhamod et al, 2020;Heydari et al, 2019;Wang, Yu, & Perdikaris, 2022). Objectives in event-based training for binary seismic event classification are only directly adversarial in the case of mislabeling, but the unknown effects of their relative importance on optimization stability and dynamics comes with an additional burden on experiment design and hyperparameter search.…”

Section: Update F θmentioning

confidence: 99%

Event‐Based Training in Label‐Limited Regimes

Linville

2022

JGR Solid Earth

View full text Add to dashboard Cite

Semi-supervised learning (SSL) offers a way to use the structural characteristics of an entire data set, regardless of label availability, for specific task learning. In this paper, structure refers to the salient waveform characteristics imparted by underlying physical processes that persist through the nonlinear transforms inherent to deep neural networks. Although an ideal feature set would exhibit structure dominated by target characteristics, in practice features can inherit dependencies which are not causally related to the prediction task. For seismic event data sets, where segmented input data spans the duration of the transient energy from a specific and discrete seismogenic process, inherited structure can come from processes including travel path, geologic site conditions, source properties, nonstationary seismic background noise, and various data artifacts. Although persistent seismic monitoring networks generate data-rich event catalogs that enable the development of deep neural network (DNN) models capable of making accurate decisions about source attributes, they do so with a complex set of features only some of which may relate to the seismogenic processes that matter for predictive modeling tasks. Yet despite imperfect feature learning, performance on seismic processing tasks using stable network geometries can be superior to traditional methods (Linville et al., 2019;Ross et al., 2019;Tibi et al., 2019) and provide capabilities that scale as data volumes increase (Nguyen et al., 2019). Therefore, even when models generalize poorly to new regions, they can add consequential value to the specific event processing pipelines they were developed for. While some authors have achieved meaningful generalization on seismic processing tasks beyond the regional networks their models were trained on (e.g., Ross et al., 2018), this work is focused on scenarios in the seismic domain where we can expect that we have collected data that imperfectly represents the set of possible states our system is capable of expressing, and we have, or can obtain at some cost, ground truth for a limited number of the observations that we have collected.

show abstract

Section: Update F θmentioning

confidence: 99%

Event‐Based Training in Label‐Limited Regimes

Linville

2022

JGR Solid Earth

View full text Add to dashboard Cite

show abstract

“…and dynamic loss reweighting strategies [143][144][145]. Instead of training all tasks together, task grouping trains only similar tasks together.…”

Section: Multi-task Learningmentioning

confidence: 99%

“…We compare the performance, in terms of test loss and energy consumption, among the following methods: 1) one by one training of activities (i.e., the vanilla multi-tenant FL); 2) all-in-one training of activities (i.e., using only activity consolidation); 3) allin-one training with gradient normalization applied to tune the gradient magnitudes among activities (GradNorm [144]); 4) estimating higher-order of activity groupings from pair-wise activities performance (HOA [147]); 5) grouping training activities with only task a nity grouping method (TAG [4]); 6) MuFL with both activity consolidation and activity splitting. Carbontracker [216] is used to measure energy consumption and carbon footprint (provided in Appendix B).…”

Section: Performance Evaluationmentioning

confidence: 99%

Toward a generic federated learning platform optimized for computer vision applications

Zhuang¹

View full text Add to dashboard Cite

Chapter 3 is published as Weiming Zhuang, Xin Gan, Yonggang Wen, and Shuai Zhang, "EasyFL: A Low-code Federated Learning Platform for Dummies", IEEE Internet of Things Journal (IOT-J), 2022. The contributions of the co-authors are as follows:• I proposed the idea, designed the system architecture, wrote the manuscript, and revised the paper.• Xin Gan and I co-implemented the system and conducted the experiments.• Prof. Yonggang Wen provided insightful comments on the idea, the system, and the manuscript.

show abstract

“…The first approach is an approximation that allows using the standard formulation of a single optimization function. This is commonly done by a linear combination of the loss functions of all tasks (SERMANET et al, 2013;MISRA et al, 2016;KOKKINOS, 2017;TEICHMANN et al, 2018;CHENNUPATI et al, 2019;SANCHEZ et al, 2019;LI et al, 2021), or some variation with adaptive weights (CIPOLLA; GAL; KENDALL, 2018;CHEN et al, 2018;JOHNS;DAVISON, 2019;LI et al, 2016;GUO et al, 2018). Although this approach is simple and has shown promising results, there are two inherent problems, as Gunantara (2018) pointed out.…”

Section: Optimization For Multi-task Learningmentioning

confidence: 99%

“…Moreover, MTL can also be used to explore the idea of creating an inductive bias between tasks (CARUANA, 1997). Due to its potential to improve generalization performance, it has gained attention in several areas of scientific and industrial communities, such as computer vision (SENER; KOLTUN, 2018;GAL;KENDALL, 2018;CHEN et al, 2018; and natural language processing (MAO et al, 2020;XUANJING, 2016;.…”

Section: Introductionmentioning

confidence: 99%

Aproveitando o comportamento de convergência para equilibrar tarefas conflitantes no aprendizado de múltiplas tarefas

Nakamura¹

View full text Add to dashboard Cite

Multi-Task Learning is a learning paradigm that uses correlated tasks to improve performance generalization. A common way to learn multiple tasks is through the hard parameter sharing approach, in which a single architecture is used to share the same subset of parameters, creating an inductive bias between them during the training process. Due to its simplicity, potential to improve generalization, and reduce computational cost, it has gained the attention of the scientific and industrial communities. In the literature, the simultaneous learning of multiple tasks is usually performed by a linear combination of loss functions. Nonetheless, tasks' gradients often conflict with each other during losses' optimization, and it is not trivial to combine them so that all tasks converge toward their optimal solution throughout the training process. To address this problem, the idea of multi-objective optimization was adopted to propose a method that takes into account the temporal behavior of the gradients to create a dynamic bias that adjusts the importance of each task during backpropagation. The result of this method is to give more attention to tasks that are diverging or not being benefited during the last iterations, ensuring that the simultaneous learning is heading to the performance maximization of all tasks. To evaluate the performance of the proposed method in learning conflicting tasks, sensitivity analysis and a series of experiments were performed on a public handwritten digit classification dataset, and on the scene understanding problem in the CityScapes Dataset. Through the performed experiments, the proposed method outperformed state-of-the-art methods in learning conflicting tasks. Unlike the adopted baselines, the proposed method ensures that all tasks reach good generalization performances at the same time it speeds up the learning curves.

show abstract

An Improved Hashing Method for Image Retrieval Based on Deep Neural Networks

Cited by 61 publications

References 18 publications

Event‐Based Training in Label‐Limited Regimes

Event‐Based Training in Label‐Limited Regimes

Toward a generic federated learning platform optimized for computer vision applications

Aproveitando o comportamento de convergência para equilibrar tarefas conflitantes no aprendizado de múltiplas tarefas

Contact Info

Product

Resources

About