Deep Active Learning via Open-Set Recognition.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Data labeling is costly and time-consuming, creating a need for efficient data selection methods.
• Active learning aims to minimize labeling costs by identifying the most informative unlabeled data points.

Purpose of the Study:

• To formulate active learning as an open-set recognition problem.
• To leverage variational neural networks (VNNs) for uncertainty estimation in active learning.
• To develop a novel active learning algorithm that handles out-of-distribution data.

Main Methods:

• Formulated active learning as an open-set recognition task.
• Utilized variational neural networks (VNNs) to measure prediction confidence (entropy).
• Selected unlabeled samples with high uncertainty (low confidence) for oracle labeling.

Main Results:

• Achieved state-of-the-art results on benchmark datasets (MNIST, CIFAR-10, CIFAR-100, FashionMNIST).
• Demonstrated robustness to out-of-distribution outliers.
• Successfully distinguished between seen and unseen datasets in unlabeled pools.

Conclusions:

• High-quality uncertainty measures are crucial for effective pool-based active learning.
• The proposed probabilistic approach offers a robust and efficient active learning strategy.
• The method shows promise for applications with limited labeling resources and noisy data.