In response to concerns over reliability and robustness in machine learning, the authors propose a multiverse analysis approach with their PRESTO framework. This framework maps the ‘multiverse’ of machine learning models that rely on latent representations, providing a better understanding of the variability in their embeddings. The authors suggest that this can help address issues of unnecessary complexity and untrustworthy representations. They demonstrate their approach theoretically and empirically, showing how it can be used to perform sensitivity analysis, detect anomalous embeddings, and navigate hyperparameter search spaces effectively.
Publication date: 5 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.01514
