D
Demerzel
Text
Deep Discrete Encoders: Identifiable Deep Generative Models with Discrete Latent Layers
A family of interpretable deep generative models, called Deep Discrete Encoders (DDEs), built as directed graphical models with several stacked binary latent layers. The central result is a set of transparent identifiability conditions guaranteeing that the model's parameters are uniquely determined by the observed-data distribution; these conditions force the latent layers to shrink progressively with depth, which both ensures consistent estimation and dictates an interpretable architecture. The paper pairs this theory with a scalable estimation pipeline -- a layerwise nonlinear spectral initialization followed by a penalized stochastic-approximation EM algorithm -- that fits models with exponentially many latent components, and validates it on hierarchical topic modeling, image representation learning, and educational response-time data.