Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Learning for Multi-class Image Classification

Thien Nhan Vo

TL;DR

This work demonstrates that pool-based active learning can substantially reduce the labeling burden for image classification without sacrificing accuracy. By formalizing four uncertainty metrics—Largest Margin, Smallest Margin, Least Confidence, and Entropy Reduction—and evaluating them on MNIST and Fruits360 with CNNs, the study reveals that smallest-margin and least-confidence generally provide the strongest data efficiency for multi-class tasks, while least-confident and entropy perform best for MNIST. The experiments show marked improvements over random sampling, especially on harder, multi-class problems, and reveal that active learning is less impactful on straightforward binary tasks. Overall, the paper provides empirical evidence that active learning is a viable, task-dependent strategy for efficient image classification in practical settings.

Abstract

A principle bottleneck in image classification is the large number of training examples needed to train a classifier. Using active learning, we can reduce the number of training examples to teach a CNN classifier by strategically selecting examples. Assigning values to image examples using different uncertainty metrics allows the model to identify and select high-value examples in a smaller training set size. We demonstrate results for digit recognition and fruit classification on the MNIST and Fruits360 data sets. We formally compare results for four different uncertainty metrics. Finally, we observe active learning is also effective on simpler (binary) classification tasks, but marked improvement from random sampling is more evident on more difficult tasks. We show active learning is a viable algorithm for image classification problems.

Active Learning for Multi-class Image Classification

TL;DR

This work demonstrates that pool-based active learning can substantially reduce the labeling burden for image classification without sacrificing accuracy. By formalizing four uncertainty metrics—Largest Margin, Smallest Margin, Least Confidence, and Entropy Reduction—and evaluating them on MNIST and Fruits360 with CNNs, the study reveals that smallest-margin and least-confidence generally provide the strongest data efficiency for multi-class tasks, while least-confident and entropy perform best for MNIST. The experiments show marked improvements over random sampling, especially on harder, multi-class problems, and reveal that active learning is less impactful on straightforward binary tasks. Overall, the paper provides empirical evidence that active learning is a viable, task-dependent strategy for efficient image classification in practical settings.

Abstract

A principle bottleneck in image classification is the large number of training examples needed to train a classifier. Using active learning, we can reduce the number of training examples to teach a CNN classifier by strategically selecting examples. Assigning values to image examples using different uncertainty metrics allows the model to identify and select high-value examples in a smaller training set size. We demonstrate results for digit recognition and fruit classification on the MNIST and Fruits360 data sets. We formally compare results for four different uncertainty metrics. Finally, we observe active learning is also effective on simpler (binary) classification tasks, but marked improvement from random sampling is more evident on more difficult tasks. We show active learning is a viable algorithm for image classification problems.
Paper Structure (17 sections, 4 equations, 20 figures, 1 algorithm)

This paper contains 17 sections, 4 equations, 20 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Active Learning
  4. Uncertainty Measures
  5. Largest Margin Uncertainty
  6. Smallest Margin Uncertainty
  7. Least Confidence Uncertainty
  8. Entropy Reduction
  9. Algorithm
  10. Experiments
  11. MNIST Dataset
  12. Fruits360 Dataset
  13. Results
  14. MNIST Results
  15. Fruits360 Results
  16. ...and 2 more sections

Figures (20)

  • Figure 1.1: Illustrative Example. Active learning enables better than random selection in classification problems.
  • Figure 6.1: MNIST Dataset of Handwritten Numbers
  • Figure 6.2: Fruits360 Dataset Classes with Examples
  • Figure 7.1: MNIST Dataset Active Training Set Accuracy Averaged Over Three Runs for Credibility
  • Figure 7.2: MNIST Dataset Active Training Set Loss Averaged Over Three Runs for Credibility
  • ...and 15 more figures