Disentangled and Distilled Encoder for Out-of-Distribution Reasoning with Rademacher Guarantees

TL;DR

This paper introduces a disentangled distilled encoder (DDE) that compresses OOD reasoning models while maintaining disentanglement, supported by Rademacher complexity guarantees, enabling efficient deployment on resource-limited devices.

Contribution

The paper proposes a novel DDE framework that combines model compression with disentanglement preservation using a formal distillation process and theoretical Rademacher guarantees.

Findings

DDE effectively reduces model size for resource-constrained deployment.

Disentanglement is preserved during distillation with theoretical guarantees.

Empirical evaluation shows maintained reasoning performance on OOD samples.

Abstract

Recently, the disentangled latent space of a variational autoencoder (VAE) has been used to reason about multi-label out-of-distribution (OOD) test samples that are derived from different distributions than training samples. Disentangled latent space means having one-to-many maps between latent dimensions and generative factors or important characteristics of an image. This paper proposes a disentangled distilled encoder (DDE) framework to decrease the OOD reasoner size for deployment on resource-constrained devices while preserving disentanglement. DDE formalizes student-teacher distillation for model compression as a constrained optimization problem while preserving disentanglement with disentanglement constraints. Theoretical guarantees for disentanglement during distillation based on Rademacher complexity are established. The approach is evaluated empirically by deploying the compressed model on an NVIDIA