Reproduce “Bundle Networks: Fiber Bundles, Local Trivializations, and a Generative Approach to Exploring Many-to-one Maps” on OpenResearch

About this paper

Nico Courts, Henry Kvinge

Bundle Networks propose a generative deep learning framework inspired by topological fiber bundles to explore the inverse mapping of many-to-one functions in machine learning. This approach accurately models and samples from the 'fibers'—all inputs corresponding to a specific output—demonstrating superior performance over state-of-the-art generative models in both fiberwise and global data generation on synthetic and real-world datasets, including a 4.9% mIoU gain on ACDC Fog.

Key insights

• Leveraging concepts from differential topology, like fiber bundles and local trivializations, provides a principled and effective way to model and explore the inverse relationships in many-to-one functions.
• Decomposing the input space into predicted labels and independent variations within local output neighborhoods is crucial for robust fiber modeling.
• Matching the topological prior for the latent fiber space to the expected fiber structure can improve performance, though the model exhibits robustness to prior choice in some real-world scenarios.

Summary by alphaXiv.