Learning Generalized Transformation Equivariant Representations Via AutoEncoding Transformations.

Area of Science:

• Computer Vision
• Machine Learning
• Representation Learning

Background:

• Convolutional Neural Networks (CNNs) rely on translation equivariance for success.
• Existing methods struggle to capture intrinsic visual structures under general transformations.
• There is a need for models that learn representations invariant to a wider range of transformations.

Purpose of the Study:

• To develop novel models for learning generalized transformation equivariant representations (GTERs).
• To capture complex visual patterns beyond linear equivariance.
• To improve performance in both unsupervised and (semi-)supervised learning tasks.

Main Methods:

• Introduced deterministic AutoEncoding Transformations (AET) and probabilistic AutoEncoding Variational Transformations (AVT).
• AET decodes transformations directly from learned representations.
• AVT maximizes mutual information between representations and transformations.

Main Results:

• The proposed models achieve generalized transformation equivariance.
• Outperformed state-of-the-art models in unsupervised and (semi-)supervised learning tasks.
• Unsupervised representations surpassed ImageNet pre-trained supervised representations in object detection.

Conclusions:

• AET and AVT effectively learn GTERs, capturing complex visual structures.
• The approach offers a powerful framework for representation learning across various tasks.
• This method advances the field of equivariant deep learning.