Learning More by Seeing Less: Structure First Learning for Efficient, Transferable, and Human-Aligned Vision

Recent advancements in computer vision have highlighted the limitations of current recognition systems, which rely heavily on rich visual inputs. In contrast, humans can interpret sparse representations, such as line drawings, with ease. The newly proposed structure-first learning paradigm leverages this insight by using line drawings as an initial training modality. This innovative approach has shown to improve model performance significantly, fostering a stronger shape bias and enhancing data efficiency across various tasks, including classification, detection, and segmentation. Notably, models trained with this method exhibit lower intrinsic dimensionality, requiring fewer principal components to capture variance, mirroring the efficient representations seen in the human brain. Furthermore, the structure-first learning paradigm enables better distillation into lightweight student models, which outperform those trained on more complex, color-supervised data. These findings not only a…