InstaDeep released ALF, an open-source modular active learning framework that runs the full data acquisition loop through five components. It offers a single API covering both offline benchmarking against existing datasets and online deployment against real-world oracles, addressing gaps in existing tools that typically support only one setting.
Researchers propose ALDA, a deployment-oriented framework that selects active learning sampling strategies for medical image classification under clinical performance constraints. ALDA fits parametric learning-curve models to candidate strategies during a short pilot phase to estimate which will meet label budgets, addressing the risk that choosing the wrong strategy can increase costs.
Researchers propose ALDA, a deployment-oriented framework that selects active learning sampling strategies for medical image classification under clinical performance constraints. ALDA fits parametric learning-curve models to candidate strategies during a short pilot phase to estimate which will meet label budgets, addressing the risk that choosing the wrong strategy can increase costs.
A new arXiv paper investigates whether open-weight LLMs can serve as standalone acquisition policies in active learning for materials discovery, evaluating five LLMs across four retrospective finite-pool tasks. The study compares LLM-driven selection against traditional probabilistic surrogate models, examining different candidate-presentation strategies.
A new arXiv paper investigates whether open-weight LLMs can serve as standalone acquisition policies in active learning for materials discovery, evaluating five LLMs across four retrospective finite-pool tasks. The study compares LLM-driven selection against traditional probabilistic surrogate models, examining different candidate-presentation strategies.
A new arXiv paper introduces a loss-based active learning approach for fine-tuning neural abstractive summarization models, aiming to reduce annotation costs by selecting only the most informative instances. The method addresses instability and computational bottlenecks in existing active learning studies for summarization.
A new arXiv paper introduces a loss-based active learning approach for fine-tuning neural abstractive summarization models, aiming to reduce annotation costs by selecting only the most informative instances. The method addresses instability and computational bottlenecks in existing active learning studies for summarization.
A new arXiv paper evaluates diversity-based active learning methods, comparing metric spaces and selection objectives like K-center and K-median. The study addresses the high cost of labeled data by exploring how to choose representative samples from unlabeled pools.
A new arXiv paper evaluates diversity-based active learning methods, comparing metric spaces and selection objectives like K-center and K-median. The study addresses the high cost of labeled data by exploring how to choose representative samples from unlabeled pools.
An arXiv paper, 'One Loop, Two Gains: Can Active Learning win the Lottery for Free?', connects the lottery ticket hypothesis with deep active learning. It observes that iterative magnitude pruning and active learning both repeatedly retrain models from scratch, raising the question of whether active learning's retraining loop can find winning tickets without additional pruning cycles.
An arXiv paper, 'One Loop, Two Gains: Can Active Learning win the Lottery for Free?', connects the lottery ticket hypothesis with deep active learning. It observes that iterative magnitude pruning and active learning both repeatedly retrain models from scratch, raising the question of whether active learning's retraining loop can find winning tickets without additional pruning cycles.