Deep Active Learning for Regression Using ε-weighted Hybrid Query Strategy

Overview

Presented at the ICML 2022 ReALML (Responsible AI and Machine Learning) workshop. Introduces an ε-weighted hybrid query strategy for deep active learning in regression tasks, combining uncertainty-based and diversity-based sampling for improved data efficiency.

Approach

The hybrid query strategy balances:

• Uncertainty sampling — selecting points where the model is least confident
• Diversity sampling — ensuring coverage of the input space
• ε-weighting — adaptive balancing between the two strategies based on training progress