Understanding the complex dynamics of large-scale biological systems requires models capable not only of extracting low-dimensional representations from noisy data, but of doing so in an interpretable fashion. This work presents a probabilistic generative framework designed for latent structure discovery in biological datasets, with a primary focus on whole-brain imaging. We integrate variational autoencoders (VAEs) with transformer autoregressive normalizing flows (TARFlow) to learn complex latent distributions over biological datasets that more accurately capture the underlying conditioning of the system. This framework provides a foundation for interpretable counterfactual explanations of statistical dependencies underlying data modalities.
Gabriela Barros (Tue,) studied this question.