DUET: 2D Structured and Approximately Equivariant Representations

TL;DR

DUET introduces structured, equivariant 2D representations that preserve transformation information, enabling better control, lower reconstruction error, and improved performance in downstream tasks compared to existing invariant or equivariant methods.

Contribution

The paper presents DUET, a novel 2D structured and equivariant representation method that maintains transformation information and enhances controllability and task performance.

Findings

DUET achieves higher accuracy on discriminative tasks.

DUET enables controlled generation with lower reconstruction error.

DUET improves transfer learning performance.

Abstract

Multiview Self-Supervised Learning (MSSL) is based on learning invariances with respect to a set of input transformations. However, invariance partially or totally removes transformation-related information from the representations, which might harm performance for specific downstream tasks that require such information. We propose 2D strUctured and EquivarianT representations (coined DUET), which are 2d representations organized in a matrix structure, and equivariant with respect to transformations acting on the input data. DUET representations maintain information about an input transformation, while remaining semantically expressive. Compared to SimCLR (Chen et al., 2020) (unstructured and invariant) and ESSL (Dangovski et al., 2022) (unstructured and equivariant), the structured and equivariant nature of DUET representations enables controlled generation with lower reconstruction error, while controllability is not possible with SimCLR or ESSL. DUET also achieves higher accuracy for several discriminative tasks, and improves transfer learning.