Support vector machines for multi-class pattern recognition

By Jason Aaron Edward Weston, Chris Watkins. 1999

How researchers see this paper

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

In Model Architectures

Citations in the corpus

It’s Not Just Size That Matters: Small Language Models Are Also Few-Shot Learners (2020, methodology): “As the training sets for RTE and CB are very small, we additionally select random unlabeled examples from MNLI (Williams et al., 2018) for both tasks.”
Improving Neural Language Modeling via Adversarial Training (2019, background): “Learning large-margin classiﬁers has been extensively studied in the literature; see e.g., Weston et al. (1999) Other Regularization Techniques for Language Models Various other techniques have been also developed to address overﬁtting in RNN language models.”
A Unified View on Multi-class Support Vector Classification (2016, methodology): “The ﬁrst all-in-one methods proposed independently by Weston and Watkins (1999), Vapnik (1998, Section 10.10), and Bredensteiner and Bennett (1999) turned out to be equivalent, up to rescaling of the decision functions and the regularization parameter C > 0.”
Online Passive-Aggressive Algorithms (2003, methodology): “This definition generalizes the definition of margin for binary classification and was employed by both single-label and multilabel learning algorithms for support vector machines (Vapnik, 1998; Weston and Watkins, 1999; Elisseeff and Weston, 2001; Crammer and Singer, 2003a).”
Image annotation using SVM (2003, background): “f w w x x ⋅ + ( ) = , ( 13 ) where wi and bi are, respectively, the normal and the bias of the hyperplane related to the i-th SVM.”
Ensemble-based classifiers (2010, methodology): “The first approach, which involves extending the algorithm, has been applied to SVMs (Weston and Watkins 1999).”
Convex Calibration Dimension for Multiclass Loss Matrices (2014, methodology): “This work generalized the framework of Bartlett et al. [2] to the multiclass 0-1 setting and used these results to study calibration of various surrogates proposed for multiclass 0-1 classiﬁcation, such as the surrogates of Weston and Watkins[30], Crammer and Singer [9], and Lee et al. [17].”
Oblivious Multi-Party Machine Learning on Trusted Processors (2016, methodology): “The SVM method is important historically and in practice for at least four separate reasons: first, it is easy to use and tune and it performs well in practice; second, it is derived from the principle of structural risk minimization [65] and comes with excellent theoretical guarantees in the form o”
A comprehensive evaluation of multicategory classification methods for microarray gene expression cancer diagnosis (2004, unknown): “This approach to multiclass SVMs is viewed by some researchers as a natural extension of the binary SVM classification problem (see Figure 2d, [Hsu2002] and [ Weston1999 ]).”
Learning Scoring Functions with Order-Preserving Losses and Standardized Supervision (2011, background): “For instance, some popular surrogate losses are based on pairwise comparisons (Weston & Watkins, 1999; Herbrich et al., 2000; Freund et al., 2003; Joachims, 2002); they are deﬁned as convex upper bounds on the pairwise disagreement , an evaluation metric which counts the number of mis-ordered pairs…”
Black-Box Tuning for Language-Model-as-a-Service (2022, methodology): “Given the output logits ˆy over a candidate set of label words, and the golden label word ˜ y of a certain sample, the cross entropy is deﬁned as For hinge loss, we adopt a multi-class extension (Weston & Watkins, 1999), In this work we set the margin γ = 2 .”
Machine Learning Paradigms for Speech Recognition: An Overview (2013, background): “One well known approach [81] is to have (31) (where sum is often replaced by max).”
Improved website fingerprinting on Tor (2013, background): “Other researchers have investigated more sophisticated techniques for multi-class classification [18].”
Large Margin Multi-Task Metric Learning (2010, background): “Support vector machines are arguably amongst the most successful classification algorithms of all times, however their multi-class extensions such as one-vs-all [4] or clever refinements of the loss functions [10, 21] all require at least one weight vector per class label.”
Good Practice in Large-Scale Learning for Image Classification (2014, background): “In [85], Weston and Watkins propose a multiclass SVM with a loss function that consists in summing the losses incurred by each class-wise score.”
Rule extraction from support vector machines (2002, unknown): “Although initially conceived for classification problems of two classes with linearly separable data, new algorithms have already been derived to solve classification problems with non-separable data [5], regression [8] and multiclass problems [2,12].”