Microscopic analysis of the latent space: heuristics for interpretability, authenticity, and bias detection in VAE representations
- Publié
- Serveur de preprints
- Zenodo
- DOI
- 10.5281/zenodo.16827724
A proposal modular heuristic framework (SBS, ABI, CLS) for microscopic analysis of latent spaces in generative models (VAEs / CVAEs). The framework combines ICA-based Uniqueness, entropic Originality and a simulated Stability heuristic to produce interpretable, multi-dimensional scores that characterise representations in the latent manifold. We validate the approach in a controlled synthetic environment and present a case study on a 10k subset of CelebA using a CVAE, showing applications for bias detection, identity-consistency analysis and discovery of high-information outliers. Code, cleaned notebook and selected outputs are provided.