Towards Making Unlabeled Data Never Hurt

By Yu-Feng Li, Zhi-Hua Zhou. 2011

352 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Representation & Self-Supervised Learning

Citations in the corpus

Concrete Problems in AI Safety (2016, background): “nce, there is a large empirical literature showing that generative approaches to semi-supervised learning based on maximizing marginal likelihood can perform very poorly when the model is misspeciﬁed [96, 108, 34, 88, 86]. The approaches discussed above all rely relatively strongly on having a well-”
A survey of data fusion in smart city applications (2019, methodology): “Research works such as Zhou et al. [261], [262] have attempted to ﬁx unlabeled data by transforming them into useful features to achieve certain objectives.”
Unsupervised Risk Estimation with only Structural Assumptions (2016, background): “However, a wide body of empirical evidence shows that, when the generative model is misspecified, the unlabeled examples can actually degrade performance (Merialdo, 1994; Cozman & Cohen, 2006; Liang & Klein, 2008; Li & Zhou, 2015).”
Teaching Semi-Supervised Classifier via Generalized Distillation (2018, background): “Cluster assumption [Joachims, 1999; Li and Zhou, 2015] holds that the examples belonging to the same class usually form a cluster in the feature space, and the true decision boundary should fall into the low density region between the well-separated clusters.”
A brief introduction to weakly supervised learning (2018, background): “This issue has been raised and studied for many years [21]; however, only recently, some solid progress are reported [52].”
Projected estimators for robust semi-supervised classification (2016, background): “An attempt at safety in semi-supervised learning was introduced in (Li & Zhou, 2011), who propose a safe variant for semi-supervised support vector machines.”
Improvability Through Semi-Supervised Learning: A Survey of Theoretical Results (2019, methodology): “An important difference between this work and the previous section is that for this paper one employs a generative model p ( x, y ) , while the SVM used by Li and Zhou (2011) is a discriminative model that inherently optimizes the class probability p ( y | x ) .”
Self-Training: A Survey (2022, neutral): “Several attempts have been made in recent years to investigate the value of unlabeled data in the training process [Castelli and Cover, 1995; Li and Zhou, 2011], and the capacity of semi-supervised learning approaches to generalize [Rigollet, 2007; Maximov et al.”
A survey on semi-supervised learning (2019, background): “Although this has received relatively little attention in the literature (likely due to publication bias, as noted by Zhu 2008), many semi-supervised learning methods only perform better than their supervised counterparts or base learners in specific cases (Li and Zhou 2015; Singh et al. 2009).”
Deformed Graph Laplacian for Semisupervised Learning (2015, methodology): “We observe that HF, S4VM, and LPDGL already achieve > 80% accuracy, which is a very encouraging result.”
Unsupervised Joint Mining of Deep Features and Image Labels for Large-Scale Radiology Image Categorization and Scene Recognition (2017, background): “Recent semi-supervised image feature learning and selftaught image recognition techniques [53, 46, 35, 27, 11] have advanced both supervised image classification and unsupervised clustering processes, demonstrating some promising results.”
How to Train Your MAML to Excel in Few-Shot Classification (2021, background): “Training a model under data budgets is important in machine learning, computer vision, and many other application fields, since the costs of collecting data [55] and labeling them [43] are by no means negligible.”
Few-shot learning with adaptively initialized task optimizer: a practical meta-learning approach (2019, background): “Considering both the data collection (Li and Zhou 2015) and labeling cost (Huang et al. 2014), it is necessary to enable a machine learning algorithm to have such a valuable ability—building classiﬁers with limited training examples, i.e., the one-shot and the few-shot Learning.”