Interpreting pretext tasks for active learning: a reinforcement learning approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural network performance scales with labeled data, but annotation is costly.
• Active learning (AL) mitigates annotation costs by selective data labeling.
• Integrating self-supervised learning (SSL) with AL presents challenges in interpreting SSL outputs for AL strategies.

Purpose of the Study:

• To propose a novel method for effectively utilizing self-supervised learning (SSL) within active learning (AL) frameworks.
• To address the uncertainty in interpreting SSL results for guiding AL.
• To enhance the efficiency and performance of deep neural network training through improved data selection.

Main Methods:

• A multi-armed bandit (MAB) approach is proposed to manage and interpret information from SSL for AL.
• A specialized data sampling process is developed to facilitate effective reinforcement learning (RL) within the AL framework.
• The method integrates SSL-derived insights into an RL-based data selection mechanism.

Main Results:

• The proposed method significantly enhances performance across multiple image classification benchmarks, including CIFAR-10, CIFAR-100, Caltech-101, SVHN, and ImageNet.
• It demonstrates superior results compared to existing active learning approaches that incorporate self-supervised learning.
• The multi-armed bandit strategy effectively leverages SSL information for more efficient data selection.

Conclusions:

• The proposed multi-armed bandit approach offers a robust solution for integrating self-supervised learning into active learning.
• This method effectively overcomes challenges in interpreting SSL outputs for active learning, leading to substantial performance gains.
• The findings suggest a promising direction for optimizing deep neural network training with limited labeled data.