Query by committee

Seung, H. Sebastian; Opper, Manfred; Sompolinsky, Haim

doi:10.1145/130385.130417

Cited by 1,268 publications

(796 citation statements)

References 6 publications

Supporting

Mentioning

757

Contrasting

Unclassified

Order By: Relevance

“…The supervised machine learning methods used in the experiments cover several different machine learning paradigms, and include Additive Regression (Friedman, 2002), Decision Table (Kohavi, 1995), Nearest Neighbor with a weighted condition (Aha et al, 1991), K* (Cleary et al, 1995), Locally Weighted Learning with Naive Bayes and Linear regression classifiers (Frank et al, 2002;Atkeson et al, 1997), Random Committee (Seung et al, 1992), and Random Trees (Aldous, 1993).…”

Section: Machine Learning Algorithmsmentioning

confidence: 99%

“…Locally Weighted Learning is a memory-based learning algorithm, used with naive Bayes or linear regression (Frank et al, 2002;Atkeson et al, 1997), and makes the prediction through the weighted connection within the data using regression or classification. The Random Committee (Seung et al, 1992) is a meta-learning algorithm that applies different randomly selected number of seeds. Random Tree (Aldous, 1993) is a tree-based method that builds a decision tree using a random selection of attributes in every given dataset.…”

Section: Machine Learning Algorithmsmentioning

confidence: 99%

See 1 more Smart Citation

Computational Estimation of Football Player Wages

Yaldo

Shamir

2017

International Journal of Computer Science in Sport

View full text Add to dashboard Cite

The wage of a football player is a function of numerous aspects such as the player's skills, performance in the previous seasons, age, trajectory of improvement, personality, and more. Based on these aspects, salaries of football players are determined through negotiation between the team management and the agents. In this study we propose an objective quantitative method to determine football players' wages based on their skills. The method is based on the application of pattern recognition algorithms to performance (e.g., scoring), behavior (e.g., aggression), and abilities (e.g., acceleration) data of football players. Experimental results using data from 6,082 players show that the Pearson correlation between the predicted and actual salary of the players is ∼0.77 (p < .001). The proposed method can be used as an assistive technology when negotiating players salaries, as well as for performing quantitative analysis of links between the salary and the performance of football players. The method is based on the performance and skills of the players, but does not take into account aspects that are not related directly to the game such as the popularity of the player among fans, predicted merchandise sales, etc, which are also factors of high impact on the salary, especially in the case of the team lead players and superstars. Analysis of player salaries in eight European football leagues show that the skills that mostly affect the salary are largely consistent across leagues, but some differences exist. Analysis of underpaid and overpaid players shows that overpaid players tend to be stronger, but are inferior in their reactions, vision, acceleration, agility, and balance compared to underpaid football players.

show abstract

Section: Machine Learning Algorithmsmentioning

confidence: 99%

Section: Machine Learning Algorithmsmentioning

confidence: 99%

Computational Estimation of Football Player Wages

Yaldo

Shamir

2017

International Journal of Computer Science in Sport

View full text Add to dashboard Cite

show abstract

“…Calculation process can vary up to selected active learning algorithm. Many algorithms are defined in literature such as query by committee [12], uncertainty sampling [13], margin sampling [14], entropy [14] etc. The most effective labeled data decided by using active learning algorithm is referred as informative data.…”

Section: Active Sample Selection In Ensemble Learningmentioning

confidence: 99%

Comparing Informative Sample Selection Strategies in Classification Ensembles

İlhan¹,

Amasyalı²

2014

IJMLC

View full text Add to dashboard Cite

Abstract-Usage of more training data with label information gives more success for classification of datasets in machine learning. But in real life, obtaining data with label information is a cost-effective and long-lasting process. Herein, active learning algorithms are emerged. Active learning algorithms aim to maintain current success rate with fewer samples in train set or increase total success of model in training process. Active learning is not only functional for regular learning methods but also can be used in ensemble learning algorithms with specified techniques. In this study, two different active learning algorithms based on class probabilities of the samples are tested on five datasets classification. Ensemble learning methods are used as classification model. Comparative results presented as graphically and numerically.

show abstract

“…A committee-based active-learning approach, called query-by-committee (QBC), was first proposed by Seung et al [7]. It was applied to selective-sampling problems by Freund [8], where the learner examined many unlabeled examples and only selected those samples that were more informative for learning than the others.…”

Section: Introductionmentioning

confidence: 99%

“…Early QBC studies by Seung et al [7] took into consideration their theoretical aspects within the context of binaryclassification problems. They defined a version space as a set of concepts that labeled all the training examples correctly, and they developed an algorithm to effectively restrict the version space as the number of examples increased.…”

Section: Introductionmentioning

confidence: 99%

Committee-Based Active Learning for Speech Recognition

Hamanaka

Sairyo

Tsutaoka

et al. 2011

IEICE Trans. Inf. & Syst.

View full text Add to dashboard Cite

SUMMARYWe propose a committee-based method of active learning for large vocabulary continuous speech recognition. Multiple recognizers are trained in this approach, and the recognition results obtained from these are used for selecting utterances. Those utterances whose recognition results differ the most among recognizers are selected and transcribed. Progressive alignment and voting entropy are used to measure the degree of disagreement among recognizers on the recognition result. Our method was evaluated by using 191-hour speech data in the Corpus of Spontaneous Japanese. It proved to be significantly better than random selection. It only required 63 h of data to achieve a word accuracy of 74%, while standard training (i.e., random selection) required 103 h of data. It also proved to be significantly better than conventional uncertainty sampling using word posterior probabilities. key words: active learning, query by committee, LVCSR, progressive alignment

show abstract

Query by committee

Cited by 1,268 publications

References 6 publications

Computational Estimation of Football Player Wages

Computational Estimation of Football Player Wages

Comparing Informative Sample Selection Strategies in Classification Ensembles

Committee-Based Active Learning for Speech Recognition

Contact Info

Product

Resources

About