[Paper Review] Learning From Less Data: Diversified Subset Selection and Active Learning in Image Classification Tasks

Vishal Kaushal, Anurag Sahoo+6 more|arXiv (Cornell University)|May 28, 2018

Machine Learning and Algorithms56 references6 citations

TL;DR

This paper proposes using diversity-driven subset selection—specifically Facility-Location and Disparity-Min models—to improve data efficiency in image classification. By selecting representative, non-redundant training subsets, the method boosts model accuracy by 2–3% with significantly less training data and reduces labeling effort in active learning, outperforming uncertainty sampling and random selection across multiple vision tasks.

ABSTRACT

Supervised machine learning based state-of-the-art computer vision techniques are in general data hungry and pose the challenges of not having adequate computing resources and of high costs involved in human labeling efforts. Training data subset selection and active learning techniques have been proposed as possible solutions to these challenges respectively. A special class of subset selection functions naturally model notions of diversity, coverage and representation and they can be used to eliminate redundancy and thus lend themselves well for training data subset selection. They can also help improve the efficiency of active learning in further reducing human labeling efforts by selecting a subset of the examples obtained using the conventional uncertainty sampling based techniques. In this work we empirically demonstrate the effectiveness of two diversity models, namely the Facility-Location and Disparity-Min models for training-data subset selection and reducing labeling effort. We do this for a variety of computer vision tasks including Gender Recognition, Scene Recognition and Object Recognition. Our results show that subset selection done in the right way can add 2-3% in accuracy on existing baselines, particularly in the case of less training data. This allows the training of complex machine learning models (like Convolutional Neural Networks) with much less training data while incurring minimal performance loss.

Motivation & Objective

• Address the high data and labeling cost of training deep CNNs in computer vision by reducing training data requirements.
• Improve active learning efficiency by minimizing redundant labeling through diversified subset selection.
• Demonstrate that diversity-aware subset selection enhances model performance, especially when training data is limited.
• Provide practical insights into selecting between Facility-Location and Disparity-Min based on dataset characteristics such as class diversity and intra-class similarity.

Proposed method

• Apply the Facility-Location function to select diverse, representative training samples that maximize coverage of the feature space.
• Use the Disparity-Min function to select the most mutually diverse samples, minimizing pairwise disparities to avoid redundancy.
• Integrate both subset selection methods with uncertainty sampling in active learning pipelines to reduce labeling effort.
• Apply greedy optimization algorithms to efficiently solve the subset selection problems, leveraging submodular properties for tractability.
• Fine-tune pre-trained models on selected subsets to evaluate performance in low-data regimes.
• Compare subset selection methods against random sampling and uncertainty sampling across multiple benchmarks: Adience, Caltech-101, MIT-67, and Gender Recognition.

Experimental results

Research questions

• RQ1 Can diversity-aware subset selection improve model accuracy when training data is limited?
• RQ2 How do Facility-Location and Disparity-Min compare in reducing labeling effort within active learning frameworks?
• RQ3 Does the performance of subset selection methods depend on dataset characteristics such as number of classes or intra-class similarity?
• RQ4 To what extent can subset selection reduce the number of human-labeled examples needed without sacrificing model accuracy?
• RQ5 In what scenarios does diversity selection outperform representative selection, and vice versa?

Key findings

• Subset selection with Facility-Location and Disparity-Min improved model accuracy by 2–3% over baseline methods, particularly in low-data regimes.
• On the Adience dataset with high intra-class similarity, Disparity-Min outperformed Facility-Location initially by selecting diverse outliers.
• On Caltech-101 and MIT-67 with more classes and higher inter-class diversity, Facility-Location performed better by selecting representative exemplars.
• In active learning, combining uncertainty sampling with subset selection consistently outperformed pure uncertainty sampling and random selection.
• As the number of unlabeled examples decreased, the advantage of subset selection diminished, indicating that uncertainty sampling alone becomes sufficient in later stages.
• The Facility-Location function serves as a strong proxy for nearest-neighbor classification, aligning with theoretical expectations from prior work.