Temporal Output Discrepancy for Loss Estimation-Based Active Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models require large annotated datasets, which are costly and time-consuming to acquire.
• Active learning strategies aim to reduce annotation costs by selecting informative samples for labeling.

Purpose of the Study:

• To develop a novel deep active learning approach that efficiently identifies informative unlabeled samples.
• To reduce the dependency on massive annotated datasets in deep learning by optimizing data selection.

Main Methods:

• Introduced a new measurement called temporal output discrepancy (TOD) to estimate sample loss.
• TOD evaluates the discrepancy of model outputs across different optimization steps to infer sample informativeness.
• Developed an unlabeled data sampling strategy and an unsupervised learning criterion based on TOD.

Main Results:

• The proposed TOD method effectively lower-bounds accumulated sample loss, enabling accurate selection of informative samples.
• Achieved superior performance compared to state-of-the-art active learning methods in image classification and semantic segmentation.
• Demonstrated the utility of TOD in selecting the best performing model from a candidate pool.

Conclusions:

• The TOD-based deep active learning approach is efficient, flexible, and task-agnostic.
• This method significantly reduces the cost and effort associated with data annotation in deep learning.
• TOD offers a robust mechanism for both sample selection and model selection in active learning.