BrAINs seminar: Learning Data Geometry: Can We Optimize Distance Graphs for Better Models? скачать в хорошем качестве

BrAINs seminar: Learning Data Geometry: Can We Optimize Distance Graphs for Better Models?

Alexander Hhvatov (ITMO) Title and Abstract of the talk: Learning Data Geometry: Can We Optimize Distance Graphs for Better Models? In this talk, I will dive into our recent work on optimizing distance graphs to better capture data geometry for specific neural network architectures and tasks. We’ll explore key questions: What makes an optimal graph? How do we choose what to optimize? By extending the classic manifold regularization problem, we explore all possible graphs to minimize Dirichlet energy, unlocking better generalization and performance. I'll compare this approach to traditional methods like Ricci flow optimization and demonstrate why it leads to superior results. Join us as we explore how understanding data geometry can drive innovation, with examples and exciting future research directions. For descriptions of the upcoming events, get the latest updates, or unsubscribe, check our Google Group https://groups.google.com/g/network-s...