Modular deep learning
Reusing, merging, and transporting knowledge across pre-trained models, plus the vision and multimodal pipelines built on top.
A good part of my research looks at deep learning models as reusable, composable parts rather than monoliths trained from scratch: merging, transporting, and adapting knowledge across pre-trained models. The same modular mindset carries into the computer vision and multimodal pipelines I build around it.