DUET: 2D Structured and Approximately Equivariant Representations
This addresses the need for more informative representations in self-supervised learning, particularly for downstream tasks requiring transformation details, though it is incremental over prior equivariant methods.
The paper tackles the problem of multiview self-supervised learning by proposing DUET, which provides 2D structured and approximately equivariant representations to retain transformation-related information, achieving lower reconstruction error and higher accuracy in discriminative tasks compared to existing methods.
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.