Self-corrected unsupervised domain adaptation

Editorial Material Computer Science, Information Systems

Using breath sound data to detect intraoperative respiratory depression in non-intubated anesthesia

Sikai Wang et al.

SCIENCE CHINA-INFORMATION SCIENCES (2021)

Article Computer Science, Information Systems

Crowdsourcing aggregation with deep Bayesian learning

Shao-Yuan Li et al.

Summary: In this study, a fully Bayesian deep generative crowdsourcing model (BayesDGC) is proposed to address a crowdsourcing classification problem, combining the strengths of deep neural networks and probabilistic graphical models. The model infers latent true labels by integrating a DNN classifier and a probabilistic model, and uses natural-gradient stochastic variational inference to tackle the inference challenge.

SCIENCE CHINA-INFORMATION SCIENCES (2021)

Add to Collection

Article Computer Science, Information Systems

Deep multiple instance selection

Xin-Chun Li et al.

Summary: This paper introduces a novel deep MIL framework, DMIS, which combines ROIs identification with deep learning. By using hard selection methods for instance selection and applying cooling down and variance normalization approaches, the framework achieves superior performance in generalization ability and positioning ROIs compared to classical MIL methods through both theoretical analysis and empirical investigations.

SCIENCE CHINA-INFORMATION SCIENCES (2021)

Add to Collection

Article Computer Science, Information Systems

On the convergence and improvement of stochastic normalized gradient descent

Shen-Yi Zhao et al.

Summary: The paper proves for the first time that stochastic normalized gradient descent (SNGD) can converge for non-convex problems, but it requires a small constant learning rate. To enhance the performance of SNGD, a new method called stagewise SNGD (S-SNGD) is proposed, which can use a larger initial learning rate and reduce it by stage.

SCIENCE CHINA-INFORMATION SCIENCES (2021)

Add to Collection

Article Computer Science, Information Systems

Learning from group supervision: the impact of supervision deficiency on multi-label learning

Miao Xu et al.

Summary: In the study of weakly supervised multi-label learning, it is common to assume one weak annotation group per instance, but recent research focuses on acquiring multiple weak annotation groups through crowdsourcing. This study introduces a new query strategy based on neural networks and analyzes factors affecting further learning outcomes.

SCIENCE CHINA-INFORMATION SCIENCES (2021)

Add to Collection

Review Computer Science, Information Systems