Interpretable Self-Supervised Learning via Representer Landmarks and Nystr\"om Approximation
Researchers have developed KREPES, a new framework designed to make self-supervised learning (SSL) models more interpretable. KREPES uses a method called Representer Landmarks, which identifies influential training examples to explain the learned representations. This framework can also quantify the transparency of these representations and has revealed biases in datasets, such as the Adult-1M dataset using demographic proxies for income. To handle large datasets like ImageNet-1K, KREPES incorporates a Nyström approximation for scalability. AI
IMPACT Enhances transparency in self-supervised models, potentially aiding bias detection and model auditing.