[Paper Review] Active Learning: Problem Settings and Recent Developments

Hideitsu Hino|arXiv (Cornell University)|Dec 8, 2020

Machine Learning and Algorithms96 references30 citations

TL;DR

This paper reviews active learning problem settings, acquisition functions, theoretical guarantees, stopping criteria, and applications, highlighting recent trends and practical considerations.

ABSTRACT

In supervised learning, acquiring labeled training data for a predictive model can be very costly, but acquiring a large amount of unlabeled data is often quite easy. Active learning is a method of obtaining predictive models with high precision at a limited cost through the adaptive selection of samples for labeling. This paper explains the basic problem settings of active learning and recent research trends. In particular, research on learning acquisition functions to select samples from the data for labeling, theoretical work on active learning algorithms, and stopping criteria for sequential data acquisition are highlighted. Application examples for material development and measurement are introduced.

Motivation & Objective

• Explain the basic problem settings of active learning and the role of labeled vs. unlabeled data in supervised learning.
• Survey acquisition functions and criteria for selecting samples to label.
• Discuss theoretical guarantees, sample/label complexity, and stopping criteria in active learning.
• Illustrate applications in measurement and material development to demonstrate practical relevance.

Proposed method

• Define the active learning framework with X (explanatory) and Y (response), hypotheses, loss functions, and version space concepts.
• Differentiate stream-based and pool-based active learning settings and discuss batch selection to improve efficiency.
• Review acquisition functions such as uncertainty, margin, entropy, BALD, and QBC, including their theoretical underpinnings.
• Explain submodularity and adaptive submodularity as tools for guarantees in batch and adaptive settings.
• Discuss learning the acquisition function via meta-learning, reinforcement learning, and bandit formulations.
• Summarize theoretical results on sample/label complexity and conditions under which active learning provides gains.

Experimental results

Research questions

• RQ1 Under what problem conditions do active learning methods reduce label complexity relative to passive learning?
• RQ2 How do acquisition functions influence the efficiency and effectiveness of learning in stream-based and pool-based settings?
• RQ3 What theoretical guarantees exist for batch and adaptive active learning, and how do submodularity concepts contribute?
• RQ4 How can acquisition functions be learned or adapted via reinforcement learning, meta-learning, or bandit frameworks?
• RQ5 What are practical considerations and stopping criteria for sequential data acquisition in real-world applications?

Key findings

• Active learning can achieve high prediction accuracy with fewer labeled samples under certain conditions and settings.
• Uncertainty-based, margin, entropy, and BALD-type acquisition functions are central to sample selection; their effectiveness depends on data distribution and model.
• Batch and adaptive submodularity provide approximation guarantees for greedy or batch selection strategies.
• Theoretical results show varying label complexity improvements depending on realizable vs. unrealizable settings and noise conditions.
• Recent work explores learning acquisition functions via reinforcement learning, meta-learning, and deep models for large-scale problems.