Epsilon Identifiability

Epsilon-identifiability addresses the challenge of recovering underlying causal relationships or latent variables from data when complete identification is impossible. Current research focuses on bounding the uncertainty in these estimations, particularly within the context of contrastive learning for multimodal data and generative models like variational autoencoders. This work aims to improve the reliability and interpretability of models by characterizing and mitigating indeterminacies, leading to more robust inferences in various scientific domains and applications. The ultimate goal is to establish conditions under which meaningful, albeit partially identified, conclusions can be drawn from limited or ambiguous data.