What is representation shift evaluation in deep neural networks?

It measures how internal feature representations change across data or model updates.

Representation shift evaluates how latent embeddings evolve between training and deployment or across model versions. Even if output metrics remain stable, internal feature drift can indicate future instability. Techniques include Centered Kernel Alignment (CKA), Singular Vector Canonical Correlation Analysis (SVCCA), and cosine similarity of embedding distributions. This is critical in debugging deep models and ensuring reproducibility.