New SSCAL method optimizes pseudo-labeling in continual active learning

A new method called SSCAL has been proposed to address the challenge of incorporating pseudo-labels into semi-supervised continual active learning (SSCAL) without increasing the overall training cost. The method formalizes the problem into two decisions: which unlabeled examples to pseudo-label and what proportion of mini-batches they should occupy. SSCAL caps the pseudo-label share and reweights labeled examples to maintain their gradient contribution, showing superior performance on average across various datasets and acquisition functions. AI

IMPACT Optimizes training efficiency for models that leverage both labeled and unlabeled data in continuous learning scenarios.

RANK_REASON Academic paper detailing a new method for semi-supervised continual active learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• SSCAL

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →