Making Your First Choice: To Address Cold Start Problem in Vision Active Learning (2210.02442v1)

Abstract: Active learning promises to improve annotation efficiency by iteratively selecting the most important data to be annotated first. However, we uncover a striking contradiction to this promise: active learning fails to select data as efficiently as random selection at the first few choices. We identify this as the cold start problem in vision active learning, caused by a biased and outlier initial query. This paper seeks to address the cold start problem by exploiting the three advantages of contrastive learning: (1) no annotation is required; (2) label diversity is ensured by pseudo-labels to mitigate bias; (3) typical data is determined by contrastive features to reduce outliers. Experiments are conducted on CIFAR-10-LT and three medical imaging datasets (i.e. Colon Pathology, Abdominal CT, and Blood Cell Microscope). Our initial query not only significantly outperforms existing active querying strategies but also surpasses random selection by a large margin. We foresee our solution to the cold start problem as a simple yet strong baseline to choose the initial query for vision active learning. Code is available: https://github.com/c-liangyu/CSVAL

• The paper proposes a novel contrastive learning solution to address biased initial queries and enhance label diversity.
• It leverages K-means clustering for pseudo-label generation, outperforming traditional and random querying strategies.
• Experimental results on CIFAR-10-LT and MedMNIST datasets show significant improvements in AUC metrics and query selection efficiency.

Addressing the Cold Start Problem in Vision Active Learning

The paper "Making Your First Choice: To Address Cold Start Problem in Vision Active Learning" addresses a critical contradiction in the field of Vision Active Learning (VAL): the common assumption that active learning delivers higher annotation efficiency at the onset of learning does not hold consistently, with random selection often proving more effective. This shortfall arises from what the authors denote as the "cold start problem," characterized by biased and outlier-prone initial queries. This work proposes an innovative solution anchored in contrastive learning to mitigate this issue by leveraging three key attributes: no need for initial annotations, pseudo-labels enhancing label diversity, and contrastive features that better identify typical data while reducing outliers.

The researchers conducted experiments on datasets such as CIFAR-10-LT and several medical imaging datasets, demonstrating that their approach significantly outperforms existing active querying strategies and random selection. Their methodology is presented as a robust baseline for initial query selection in VAL, potentially setting new standards for vision tasks.

Problem Framing and Methodological Insights

Active learning strives to improve model performance through iterative selection of data points for annotation, but this process is impeded during initial cycles by biased class representation and the inclusion of outliers. The paper identifies two prime causes: biased queries favoring certain classes and reliance on models inadequately equipped to handle outliers at the start of the learning process. Addressing these issues, the authors leverage contrastive learning and clustering via K-means to enhance initial query diversity and sample typical data points.

Contrastive Learning Approach:

1. Label Diversity: The use of pseudo-labels generated by K-means clustering ensures label diversity, achieving significant gains over traditional methods biased towards majority classes.
2. Typical Data Identification: Through instance discrimination, the framework identifies data challenging to distinguish (i.e., hard-to-contrast data), reducing the likelihood of selecting outliers early in the learning cycle.

Experimental Results and Validations

The experiments conducted on CIFAR-10-LT and various MedMNIST datasets (PathMNIST, BloodMNIST, OrganAMNIST) provided compelling evidence of the approach's efficiency. The proposed strategy not only covered a wider class distribution but also significantly outperformed traditional and random methods in terms of Area Under Curve (AUC) metrics.

Significant improvements were noted:

• On CIFAR-10-LT, the solution exceeded random selection by 21.2% to 24.1% in AUC scores at higher querying budgets.
• On PathMNIST, OrganAMNIST, and BloodMNIST, the gains against random selection ranged from 1.8% to 5.2% at minimal budget allocations.

Implications and Future Directions

The method's efficacy in addressing the cold start problem positions it as a foundational strategy for initial query selection in active learning for image classification and potentially broader computer vision applications. By ensuring a diverse class representation and focusing on typical data, the approach paves the way for more efficient learning cycles in scenarios with limited labeled data.

Broader Implications:

• Theoretical Impact: This paper reinforces the importance of initial query selection, challenging previous assumptions about the superiority of active querying over random sampling and establishing a theoretical basis for the use of self-supervised learning strategies.
• Practical Deployment: The framework's application, especially in highly imbalanced domains like medical imaging, suggests a pathway to reduce annotation overhead and improve learning efficiency in high-stakes environments.

Future Directions:

• Adaptive Strategies: Further exploration into adaptive budget allocation strategies that dynamically adjust based on the querying phase could be worthwhile.
• Domain-Specific Adaptations: Applying and evaluating the approach in different fields, including robotics and autonomous systems, could provide insights on necessary adaptations for domain-specific peculiarities.
• Integration with Other Active Learning Frameworks: The integration of this methodology with unsupervised or semi-supervised approaches might yield synergistic benefits, enhancing the baseline further.

This paper contributes substantially to active learning, providing a strategic pivot from conventional methodologies and opening avenues for future inquiry and optimization.