Do sampling-based neural models exhibit marginalization and change of variables in practice?

Demonstrate that sampling-based neural representations (which treat neural variability as posterior samples) actually perform marginalization and change-of-variables computations, as evidenced either by behavioral generalization patterns or by measurable downstream neuronal effects consistent with these probabilistic operations.

Background

Sampling models naturally afford marginalization and change of variables in principle: expectations can be computed by averaging samples, and transformations can be applied directly to samples. However, the paper notes a lack of empirical or modeling demonstrations that such operations are indeed executed in behavior or in downstream neural readouts.

Establishing these operations would provide critical support that sampling-based codes are used probabilistically in the sense required by the paper’s criteria for probabilistic transfer.