Motivated by pervasive role of orthogonality in natural data and models, we propose orthogonality as a cornerstone for principled neural network development:

To directly investigate the effect of intra-data orthogonality, we applied the orthogonality principle to neural networks tasked with super-resolution analysis of fluid dynamics, where intrinsic orthogonality exists among spatial velocity components.

In the inter-data orthogonality context, we explored the efficacy of orthogonality by employing the CLIP model as the baseline architecture. By modeling orthogonal relationships between different modalities of CLIP, the recognition performance can obtain obvious improvements.

By incorporating orthogonality within the architecture of the neural network, we evaluate the effectiveness of orthogonality for parameter-efficient fine-tuning of the SAM model. The integration of the orthogonality principle facilitates the modeling of relationships among parameters from different layers in an orthogonal manner, thereby validating its effectiveness in intra-model orthogonality scenarios.

To further validate the effectiveness of the orthogonality principle in the context of inter-model orthogonality, we examine the transferability of the natural-vision-based CLIP model across diverse domains, focusing on remote sensing and medical imaging with domain-specific models. By establishing inter-model orthogonal relationships between these distinct domains, we observe significant performance improvements, demonstrating the utility of orthogonality in enhancing cross-domain adaptability and efficacy.