Deep probabilistic model synthesis enables unified modeling of whole-brain neural activity across individual subjects

A team of researchers including William E. Bishop and James E. Fitzgerald has introduced a novel machine learning framework called Deep Probabilistic Model Synthesis (DPMS). Published on arXiv, the method addresses a critical challenge in fields like neuroscience: how to synthesize quantitative models from data collected across multiple instances of the same general system, such as the brains of different individual animals. Traditional ML models typically analyze one instance at a time, limiting the ability to derive general principles. DPMS leverages variational inference to learn a shared conditional prior distribution over model parameters, which ties together all system instances, while also learning instance-specific posterior distributions to capture the unique structure of each individual.

DPMS is designed to be versatile, capable of synthesizing a wide array of model classes including those for regression, classification, and dimensionality reduction. The researchers validated their framework on both synthetic data and real-world whole-brain neural activity data from larval zebrafish. In these demonstrations, DPMS showed a measurable improvement over models trained on single instances alone, proving its efficacy in creating a more robust, unified understanding from disparate datasets. This approach moves beyond treating each subject's data in isolation, enabling the construction of a generalized model that still accounts for individual variation.