Canonicalization Network

Canonicalization networks aim to transform diverse input data into a standardized, or "canonical," representation, facilitating downstream tasks by mitigating variations irrelevant to the core information. Current research focuses on improving the efficiency and accuracy of these networks, particularly for large pretrained models and applications involving symmetries like rotations, often employing contrastive learning or dataset-dependent priors to guide the canonicalization process. This approach enhances the robustness and efficiency of various machine learning models, impacting fields ranging from point cloud processing and motion retargeting to knowledge base construction and data cleaning by enabling the use of simpler, more efficient models while maintaining or improving performance.