ProReGen: Progressive Residual Generation under Attribute Correlations

Area of Science:

• Computer Vision
• Machine Learning
• Deep Generative Models

Background:

• Attribute correlations in training data compromise deep generative models' (DGMs) ability to synthesize under-represented attribute combinations (minority samples).
• Existing methods like data re-sampling, pseudo-supervision, or inductive bias have limitations in addressing this challenge.

Purpose of the Study:

• To introduce ProReGen (progressive residual generation), a novel approach inspired by Robinson's transformation, to enhance the synthesis of minority samples by DGMs.
• To address the challenge of learning DGMs conditioned on correlated inputs by transforming the problem to learning conditioned on orthogonal inputs.

Main Methods:

• ProReGen partially outtakes predictable attribute components, creating a residual attribute that is independent of other image attributes.
• The model progressively learns the generation by first utilizing abundant majority samples and then refining with limited minority samples using residual attributes.
• Employs input orthogonalization and progressive residual learning inspired by Robinson's transformation.

Main Results:

• ProReGen transforms the learning of DGMs from correlated to orthogonal inputs, simplifying the generation process.
• The progressive learning strategy effectively utilizes both majority and minority samples for improved generation.
• Demonstrated improved correctness of minority generations on benchmark datasets with varying attribute correlations compared to existing strategies.

Conclusions:

• ProReGen offers a significant improvement in synthesizing minority samples for DGMs, overcoming limitations of existing methods.
• The method's effectiveness is validated across diverse datasets, highlighting its robustness in handling attribute correlations.
• ProReGen provides a promising direction for developing more equitable and accurate generative models.