Universal Geometric Transfer Learning

In this project, we propose to develop a theoretical and practical framework for transfer learning with geometric 3D data. Most existing learning-based approaches, aimed at analyzing 3D data, are based on training neural networks from scratch for each data modality and application. This means that such methods, first, ignore the wider information overlap that might exist across different tasks and object or scene categories, and, second, tend to generalize poorly beyond the specific scenarios for which they are trained. Even more fundamentally, the majority of existing techniques are limited to problem settings in which sufficient amount of training data is available, making them ill-adapted in many practical applications with limited supervision.In this project, we suggest to take a fundamentally different approach to geometric data analysis: rather than designing independent application or class-specific solutions, we propose to develop a theoretical and practical framework for geometric transfer learning. Our main goal will be to develop universally-applicable methods by combining powerful pre-trainable modules with effective multi-scale analysis and fine-tuning, given minimal task-specific data. The overall key to our study will be analyzing rigorous ways, both theoretically and in practice, in which solutions can be transferred and adapted across problems, semantic categories and geometric data types.Such an approach will open the door to fundamentally new tasks and modeling tools, applicable to any geometric data analysis scenario, regardless of the amount of training data available. This would allow, for example, to track the evolution of biological systems, by studying the underlying complex 3D shape dynamics, or to analyze variability in object and scene collections consisting of 3D scans of previously unseen shape categories, crucial in cultural preservation and life science applications, among myriad others.