Scalable active learning for multiclass image classification

Ajay J. Joshi, Fatih Porikli, Nikolaos P Papanikolopoulos

Research output: Contribution to journalArticle

73 Citations (Scopus)

Abstract

Machine learning techniques for computer vision applications like object recognition, scene classification, etc., require a large number of training samples for satisfactory performance. Especially when classification is to be performed over many categories, providing enough training samples for each category is infeasible. This paper describes new ideas in multiclass active learning to deal with the training bottleneck, making it easier to train large multiclass image classification systems. First, we propose a new interaction modality for training which requires only yes-no type binary feedback instead of a precise category label. The modality is especially powerful in the presence of hundreds of categories. For the proposed modality, we develop a Value-of-Information (VOI) algorithm that chooses informative queries while also considering user annotation cost. Second, we propose an active selection measure that works with many categories and is extremely fast to compute. This measure is employed to perform a fast seed search before computing VOI, resulting in an algorithm that scales linearly with dataset size. Third, we use locality sensitive hashing to provide a very fast approximation to active learning, which gives sublinear time scaling, allowing application to very large datasets. The approximation provides up to two orders of magnitude speedups with little loss in accuracy. Thorough empirical evaluation of classification accuracy, noise sensitivity, imbalanced data, and computational performance on a diverse set of image datasets demonstrates the strengths of the proposed algorithms.

Original languageEnglish (US)
Article number6127880
Pages (from-to)2259-2273
Number of pages15
JournalIEEE Transactions on Pattern Analysis and Machine Intelligence
Volume34
Issue number11
DOIs
StatePublished - Oct 1 2012

Fingerprint

Multi-class Classification
Active Learning
Image classification
Image Classification
Modality
Value of Information
Training Samples
Object recognition
Computer vision
Seed
Learning systems
Labels
Hashing
Feedback
Object Recognition
Multi-class
Approximation
Locality
Large Data Sets
Computer Vision

Keywords

  • Active learning
  • multiclass classification
  • object recognition
  • scalable machine learning

Cite this

Scalable active learning for multiclass image classification. / Joshi, Ajay J.; Porikli, Fatih; Papanikolopoulos, Nikolaos P.

In: IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 34, No. 11, 6127880, 01.10.2012, p. 2259-2273.

Research output: Contribution to journalArticle

@article{ef40859ced554083b91c49157e14397b,
title = "Scalable active learning for multiclass image classification",
abstract = "Machine learning techniques for computer vision applications like object recognition, scene classification, etc., require a large number of training samples for satisfactory performance. Especially when classification is to be performed over many categories, providing enough training samples for each category is infeasible. This paper describes new ideas in multiclass active learning to deal with the training bottleneck, making it easier to train large multiclass image classification systems. First, we propose a new interaction modality for training which requires only yes-no type binary feedback instead of a precise category label. The modality is especially powerful in the presence of hundreds of categories. For the proposed modality, we develop a Value-of-Information (VOI) algorithm that chooses informative queries while also considering user annotation cost. Second, we propose an active selection measure that works with many categories and is extremely fast to compute. This measure is employed to perform a fast seed search before computing VOI, resulting in an algorithm that scales linearly with dataset size. Third, we use locality sensitive hashing to provide a very fast approximation to active learning, which gives sublinear time scaling, allowing application to very large datasets. The approximation provides up to two orders of magnitude speedups with little loss in accuracy. Thorough empirical evaluation of classification accuracy, noise sensitivity, imbalanced data, and computational performance on a diverse set of image datasets demonstrates the strengths of the proposed algorithms.",
keywords = "Active learning, multiclass classification, object recognition, scalable machine learning",
author = "Joshi, {Ajay J.} and Fatih Porikli and Papanikolopoulos, {Nikolaos P}",
year = "2012",
month = "10",
day = "1",
doi = "10.1109/TPAMI.2012.21",
language = "English (US)",
volume = "34",
pages = "2259--2273",
journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
issn = "0162-8828",
publisher = "IEEE Computer Society",
number = "11",

}

TY - JOUR

T1 - Scalable active learning for multiclass image classification

AU - Joshi, Ajay J.

AU - Porikli, Fatih

AU - Papanikolopoulos, Nikolaos P

PY - 2012/10/1

Y1 - 2012/10/1

N2 - Machine learning techniques for computer vision applications like object recognition, scene classification, etc., require a large number of training samples for satisfactory performance. Especially when classification is to be performed over many categories, providing enough training samples for each category is infeasible. This paper describes new ideas in multiclass active learning to deal with the training bottleneck, making it easier to train large multiclass image classification systems. First, we propose a new interaction modality for training which requires only yes-no type binary feedback instead of a precise category label. The modality is especially powerful in the presence of hundreds of categories. For the proposed modality, we develop a Value-of-Information (VOI) algorithm that chooses informative queries while also considering user annotation cost. Second, we propose an active selection measure that works with many categories and is extremely fast to compute. This measure is employed to perform a fast seed search before computing VOI, resulting in an algorithm that scales linearly with dataset size. Third, we use locality sensitive hashing to provide a very fast approximation to active learning, which gives sublinear time scaling, allowing application to very large datasets. The approximation provides up to two orders of magnitude speedups with little loss in accuracy. Thorough empirical evaluation of classification accuracy, noise sensitivity, imbalanced data, and computational performance on a diverse set of image datasets demonstrates the strengths of the proposed algorithms.

AB - Machine learning techniques for computer vision applications like object recognition, scene classification, etc., require a large number of training samples for satisfactory performance. Especially when classification is to be performed over many categories, providing enough training samples for each category is infeasible. This paper describes new ideas in multiclass active learning to deal with the training bottleneck, making it easier to train large multiclass image classification systems. First, we propose a new interaction modality for training which requires only yes-no type binary feedback instead of a precise category label. The modality is especially powerful in the presence of hundreds of categories. For the proposed modality, we develop a Value-of-Information (VOI) algorithm that chooses informative queries while also considering user annotation cost. Second, we propose an active selection measure that works with many categories and is extremely fast to compute. This measure is employed to perform a fast seed search before computing VOI, resulting in an algorithm that scales linearly with dataset size. Third, we use locality sensitive hashing to provide a very fast approximation to active learning, which gives sublinear time scaling, allowing application to very large datasets. The approximation provides up to two orders of magnitude speedups with little loss in accuracy. Thorough empirical evaluation of classification accuracy, noise sensitivity, imbalanced data, and computational performance on a diverse set of image datasets demonstrates the strengths of the proposed algorithms.

KW - Active learning

KW - multiclass classification

KW - object recognition

KW - scalable machine learning

UR - http://www.scopus.com/inward/record.url?scp=84866702739&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84866702739&partnerID=8YFLogxK

U2 - 10.1109/TPAMI.2012.21

DO - 10.1109/TPAMI.2012.21

M3 - Article

VL - 34

SP - 2259

EP - 2273

JO - IEEE Transactions on Pattern Analysis and Machine Intelligence

JF - IEEE Transactions on Pattern Analysis and Machine Intelligence

SN - 0162-8828

IS - 11

M1 - 6127880

ER -