Unsupervised feature learning for self-tuning neural networks.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning enables model adaptation across domains, but typically requires labeled data.
• Fine-tuning is a common transfer learning technique, yet its reliance on labeled target data limits applicability.
• Existing methods often necessitate labeled datasets, restricting transfer learning to new domains.

Purpose of the Study:

• To propose a fully unsupervised self-tuning algorithm for learning visual features across different domains.
• To overcome the limitations of labeled data requirements in traditional transfer learning methods.
• To enhance the adaptability of pre-trained models to novel target domains without supervision.

Main Methods:

• The proposed method updates pre-trained models by minimizing triplet loss using unlabeled target domain data.
• A relevance measure for unlabeled data is introduced using a bagged clustering approach.
• Triplets (anchor, positive, negative) are sampled based on relevance scores and Euclidean distances in feature space.

Main Results:

• The fully unsupervised self-tuning algorithm significantly improves network performance.
• Extensive evaluations on five benchmark datasets demonstrate enhanced classification accuracy, feature analysis, and clustering quality.
• Applying the self-tuning method to fine-tuned networks yields superior results.

Conclusions:

• The proposed unsupervised self-tuning algorithm effectively learns visual features in new domains without labeled data.
• This approach broadens the applicability of transfer learning, especially in data-scarce scenarios.
• The method offers a significant advancement in unsupervised domain adaptation for visual learning.