DPMS pools biological models without matching every neuron

Researchers at Janelia, Radboud University, and Northwestern University have introduced Deep Probabilistic Model Synthesis, or DPMS, a hierarchical variational-inference framework for learning across subjects whose internal components do not align. The DPMS paper evaluates the method on synthetic benchmarks and whole-brain calcium imaging recordings from larval zebrafish, with roughly 80,000 neurons recorded per fish.

DPMS links subject-specific models through a shared prior learned from component metadata, such as each neuron’s three-dimensional position. Data-rich subjects can therefore inform models for subjects with fewer observations while each subject retains its own parameter distribution.

Neuron IDs break naive pooling

Standard pooled training assumes that a feature has the same meaning in every sample. Biological recordings violate that assumption: neuron 4,217 in one fish has no direct counterpart in another, and even similar cells may occupy different positions or respond differently.

Fitting each subject independently preserves its structure. Those separate estimates cannot easily share evidence, and averaging their parameters is poorly defined when components lack one-to-one correspondence. The same problem appears in patient-specific physiological models, device-specific sensor models, and user-specific recommendation systems.

Coordinates become a shared prior

DPMS casts the collection of subject models as a hierarchy. A Conditional Prior Distribution, or CPD, maps auxiliary properties for each component to a probability distribution over that component’s model parameters. In the zebrafish experiments, three-dimensional neuron coordinates provide the conditioning information; other applications could use cell types, anatomical labels, or functional measurements.