Advancements in Active Learning and Foundation Models

The field of machine learning is witnessing significant developments in active learning and foundation models. Researchers are exploring innovative methods to combat the low data problem, including data augmentation, semi-supervised learning, and active learning. While active learning is found to be less efficient than other methods in certain scenarios, its combination with strong data augmentation and semi-supervised learning techniques can still provide improvements. Foundation models are being investigated for their potential in medical image analysis, with pre-trained models showing powerful feature extraction abilities. The use of foundation models for mitotic figure classification has demonstrated superior performance, reaching levels close to 100% data availability with only 10% of training data. Noteworthy papers in this area include: MedCAL-Bench, which proposes a comprehensive benchmark for cold-start active learning with foundation models for medical image analysis. AdaFusion, which introduces a novel prompt-guided inference framework that dynamically integrates complementary knowledge from multiple pathology foundation models. ALScope, which presents a unified toolkit for deep active learning, integrating multiple datasets and algorithms for comprehensive evaluation. Benchmarking Foundation Models for Mitotic Figure Classification, which investigates the use of foundation models for mitotic figure classification and evaluates their robustness to unseen tumor domains.

Sources

The Role of Active Learning in Modern Machine Learning

MedCAL-Bench: A Comprehensive Benchmark on Cold-Start Active Learning with Foundation Models for Medical Image Analysis

Active Learning and Transfer Learning for Anomaly Detection in Time-Series Data

Benchmarking Foundation Models for Mitotic Figure Classification

ALScope: A Unified Toolkit for Deep Active Learning

AdaFusion: Prompt-Guided Inference with Adaptive Fusion of Pathology Foundation Models

Cold Start Active Preference Learning in Socio-Economic Domains

Built with on top of