Interpretation of recurrent DNNs that explain reaction-time variance

Ascertain whether recurrent deep neural network models that account for image-by-image reaction time variance in object recognition should be interpreted as generative models of inference or as discriminative models, and determine whether improving such models shifts their interpretation toward one framework or the other.

Background

Adding recurrence to deep networks improves alignment with human reaction-time variability, but it is uncertain whether these mechanistic models correspond to generative or discriminative computations at the algorithmic level.

The authors note that clarifying this interpretational status—and how it might evolve as models are improved—remains unresolved.