Label Embedding Trees for Large Multi-Class Tasks

By Samy Bengio, Jason Aaron Edward Weston, David Grangier. Neural Information Processing Systems, 2010

421 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/a3ea706f6604a1e6e87c33d7a3b4b97b1bb338ef

How researchers see this paper

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

Citations in the corpus

30Hz Object Detection with DPM V5 (2014, background): “[24] use detectors that are suitable to discriminate between groups of categories.”
Dynamic Deep Neural Networks: Optimizing Accuracy-Efficiency Trade-offs by Selective Execution (2017, methodology): “Input-dependent execution has been widely used in computer vision, from cascaded detectors [31, 15] to hierarchical classification [10, 6].”
Recursive regularization for large-scale classification with hierarchical and graphical dependencies (2013, background): “The key advantage of HR-{SVM,LR} over other hierarchical models such as [31], [6], [4], [40] is that there are no constraints that maximizes the margin between correct and incorrect predictions.”
Learning fine-grained features via a CNN Tree for Large-scale Classification (2015, methodology): “[47] propose a label embedding tree for large-scale classification.”
Nonlinear Mixup: Out-Of-Manifold Data Augmentation for Text Classification (2020, methodology): “To cope with this challenge, we adopt the idea of label embedding (Bengio, Weston, and Grangier 2010).”
Multi-label learning with millions of labels: recommending advertiser bid phrases for web pages (2013, methodology): “Recent innovative work has focussed on learning hierarchies [3, 4, 18] for multi-class problems with”
Attribute-Based Detection of Unfamiliar Classes with Humans in the Loop (2013, background): “Approaches to large-scale multi-class classification address the computational infeasibility of testing against all possible classes by using a hierarchical tree structure of classifiers [15] [1] [8].”
Scaling Graph-based Semi Supervised Learning to Large Number of Labels Using Count-Min Sketch (2013, background): “Other approaches to scaling classifiers to many classes include reducing the dimension of label space, for instance by error-correcting output codes [9] or label embedding trees [3].”
DeViSE: A Deep Visual-Semantic Embedding Model (2013, methodology): “This last step can be done efficiently using either a tree or a hashing technique, in order to be faster than the naive linear search approach (see for instance [1]).”
MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space (2021, background): “Some works aim to learn a label tree structure when taxonomy is unavailable [2, 9, 30].”
Sequential Max-Margin Event Detectors (2014, methodology): “[2] applied tree-traversing to reduce the computational cost of multi-class activity classification in video.”
iPrivacy: Image Privacy Protection by Identifying Sensitive Objects via Deep Multi-Task Learning (2017, background): “I. INTRODUCTION W ITH the growing popularity of smart-phones and othermobile devices, high-quality cameras are increasingly pervasive.”
Multi-Task Label Embedding for Text Classification (2017, background): “The idea of embedding label is not new and is primarily designed to improve general neural text classification models (Bengio et al., 2010; Norouzi et al., 2013; Ma et al., 2016), but here we mainly focus on integrating label embedding to enhance MTL.”
Hierarchical Regularization Cascade for Joint Learning (2013, background): “The multitask loss is defined as follows: L({Si}ki=1,W) = k∑ i=1 ∑ s∈Si max(0, 1− ys ∗ f i(xs)) (1) Without joint regularization this is just the sum of the loss of k individual tasks.”
Large Scale Visual Recognition (2012, methodology): “The majority of existing models are tree based, including Filter Tree [42], Conditional Probability Tree(CPT) [43], and Label Trees [26].”
Beyond One-hot Encoding: lower dimensional target embedding (2018, background): “Due to their computing efﬁciency based on a divide-and-conquer strategy, output embeddings have been also proven useful for those multi-class classiﬁcation problems in which testing all possible class labels and hierarchical structures is not feasible [4, 41, 43, 7].”
Meta-class features for large-scale object categorization on a budget (2012, methodology): “Our method is an instance of the algorithm for label tree learning described in [2].”
Joint Feature Selection and Subspace Learning for Cross-Modal Retrieval (2016, background): “There is also some related work on multi-view embedding methods [46], [47], [48], [49], [50], which have been applied to document categorization, image annotation, multi-class classification, etc.”