A Survey on Deep Active Learning: Recent Advances and New Frontiers

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Active learning enhances model performance with fewer training samples by iteratively querying an oracle for labels.
• Deep active learning (DAL) is a rapidly growing subfield with limited comprehensive surveys available.
• Existing literature lacks a structured overview of DAL methodologies and applications.

Purpose of the Study:

• To provide an advanced and comprehensive survey of deep active learning (DAL).
• To systematically categorize and analyze existing DAL methods.
• To guide researchers in navigating challenges and identifying future research avenues in DAL.

Main Methods:

• Conducted a thorough review and filtering of relevant academic papers.
• Formally defined the DAL task, outlining key baselines and datasets.
• Developed a taxonomy of DAL methods based on annotation types, query strategies, model architectures, learning paradigms, and training processes.

Main Results:

• Presented a systematic taxonomy of DAL techniques, analyzing their respective strengths and weaknesses.
• Summarized the primary applications of DAL across Natural Language Processing (NLP), Computer Vision (CV), and Data Mining (DM).
• Identified current challenges and outlined future research perspectives in the field of DAL.

Conclusions:

• This survey serves as a crucial resource for researchers entering the field of deep active learning.
• It provides a structured understanding of DAL methodologies, facilitating further advancements.
• The work aims to accelerate progress in the burgeoning area of efficient machine learning model training.