“…Partial label learning (PLL), also known as superset-label learning Dietterich 2012, 2014) To tackle the mentioned challenge, existing works mainly focus on disambiguation (Feng and An 2019;Nguyen and Caruana 2008;Zhang and Yu 2015;Wang, Zhang, and Li 2022;Fan et al 2021;Xu, Lv, and Geng 2019;Zhang, Wu, and Bao 2022;Qian et al 2023), which can be broadly divided into two categories: averaging-based approaches and identification-based approaches. For the averaging-based approaches (Hüllermeier and Beringer 2005;Cour, Sapp, and Taskar 2011;Zhang and Yu 2015), each candidate label of a training sample is treated equally as the ground-truth one and the final prediction is yielded by averaging the modeling outputs.…”
In partial label learning (PLL), each instance is associated with a set of candidate labels among which only one is ground-truth. The majority of the existing works focuses on constructing robust classifiers to estimate the labeling confidence of candidate labels in order to identify the correct one. However, these methods usually struggle to rectify mislabeled samples. To help existing PLL methods identify and rectify mislabeled samples, in this paper, we introduce a novel partner classifier and propose a novel ``mutual supervision'' paradigm. Specifically, we instantiate the partner classifier predicated on the implicit fact that non-candidate labels of a sample should not be assigned to it, which is inherently accurate and has not been fully investigated in PLL. Furthermore, a novel collaborative term is formulated to link the base classifier and the partner one. During each stage of mutual supervision, both classifiers will blur each other's predictions through a blurring mechanism to prevent overconfidence in a specific label. Extensive experiments demonstrate that the performance and disambiguation ability of several well-established stand-alone and deep-learning based PLL approaches can be significantly improved by coupling with this learning paradigm.
“…Partial label learning (PLL), also known as superset-label learning Dietterich 2012, 2014) To tackle the mentioned challenge, existing works mainly focus on disambiguation (Feng and An 2019;Nguyen and Caruana 2008;Zhang and Yu 2015;Wang, Zhang, and Li 2022;Fan et al 2021;Xu, Lv, and Geng 2019;Zhang, Wu, and Bao 2022;Qian et al 2023), which can be broadly divided into two categories: averaging-based approaches and identification-based approaches. For the averaging-based approaches (Hüllermeier and Beringer 2005;Cour, Sapp, and Taskar 2011;Zhang and Yu 2015), each candidate label of a training sample is treated equally as the ground-truth one and the final prediction is yielded by averaging the modeling outputs.…”
In partial label learning (PLL), each instance is associated with a set of candidate labels among which only one is ground-truth. The majority of the existing works focuses on constructing robust classifiers to estimate the labeling confidence of candidate labels in order to identify the correct one. However, these methods usually struggle to rectify mislabeled samples. To help existing PLL methods identify and rectify mislabeled samples, in this paper, we introduce a novel partner classifier and propose a novel ``mutual supervision'' paradigm. Specifically, we instantiate the partner classifier predicated on the implicit fact that non-candidate labels of a sample should not be assigned to it, which is inherently accurate and has not been fully investigated in PLL. Furthermore, a novel collaborative term is formulated to link the base classifier and the partner one. During each stage of mutual supervision, both classifiers will blur each other's predictions through a blurring mechanism to prevent overconfidence in a specific label. Extensive experiments demonstrate that the performance and disambiguation ability of several well-established stand-alone and deep-learning based PLL approaches can be significantly improved by coupling with this learning paradigm.
“…The video question answering task associates high dimensional samples (videos and questions) and low dimensional labels (answers). Supervised strategies give rise to the possibility that samples could be associated with more than one label, or are improperly labelled [32,33]. Unsupervised learning therefore has an opportunity to shine, with the capability to reduce the sample dimension or identify clusters.…”
Anomaly detection by tracking if the context of a video stream has changed could be useful, but supervised training to classify video context can be cumbersome and error prone. Instead, we apply a cascade of clustering techniques that operate on a weakly supervised video data lake to extract a context representation of a video sequence. We then train a bi-directional LSTM model to mimic the functionality of the cascade and predict a context representation from video. Additional experiments have shown that if the context is fed as an additional input to a legacy Video Question Answering solution, loss improves by more than 20% relative to it's baseline after training over 120 epochs, which is significant as current state of the art accuracy for VideoQA solutions is close to 50%. This report is also a demonstration of how to chart a path to freedom from the requirement to explicitly label data, while preserving semantics.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.