A Unifying Framework for Concept-Based Representational Similarity
This work proposes a unified framework to understand concept alignment by decomposing it into four properties: instance-wise and distributional variants of translation and concept consistency. It shows that optimizing one property, like translation, does not guarantee others. The Coupled Sparse Autoencoder (CoSAE) demonstrates that jointly enforcing these multi-objective properties leads to superior performance compared to existing methods.