The Significance of Functional View in Score-Based Generative Modeling
Score-based models have become popular in generating images, text, and even molecules. However, designing the score network to adapt to different data domains can be challenging. At the Diffusion Models workshop at NeurIPS 2023, a paper was accepted that tackles this problem by taking a functional view of data.
The Functional View
By taking a functional view, it becomes possible to create a common shared representation for seemingly different data domains. This allows for the re-formulation of the score function to handle functional data, making it applicable to various modalities such as images, geometry, and video.
Application to Different Modalities
This unified architecture has shown effective application to different modalities including images, geometry, and video. It has also demonstrated the ability to learn generative models of signals defined on non-euclidean geometry.
Conclusion
By adopting a functional view in score-based generative modeling, it becomes possible to address the challenge of designing the score network for different data domains. This opens up new possibilities for the application of score-based models in various domains, making it a significant advancement in the field of AI.