David Pfau: Manifold Factorization and AI for Science

Spectral Methods

Spectral methodologies in machine learning are explored through the lens of optimization and eigen decomposition. David Pfau discusses the challenges of translating spectral learning into the deep learning era, highlighting the difficulty of viewing spectral decompositions as optimization problems 1. He emphasizes the importance of understanding matrix manifolds and nonlinear dimensionality reduction, which were influential in his early career 2. Pfau notes, "I've never seen anybody do that before. And that was what really excited me about this and what really got me into this" 3.

   

Optimization Challenges

The optimization of spectral learning models presents unique challenges, particularly in non-convex settings. David Pfau explains the convergence of power iteration despite its non-convex nature, which defies conventional optimization intuitions 4. He reflects on the ongoing difficulties in solving adversarial problems in vision, noting that while progress has been made, there are still threads to pull on 5. Pfau describes the beauty of theoretical work, stating, "It has both this level of mathematical precision and also there's sort of a simple but universal mechanism that implies" 6.

   

Spectral Applications

Spectral inference networks find applications in various fields, notably in physics for understanding complex phenomena. David Pfau highlights the connection between spectral methods and slow feature analysis, which appeals to theoretical neuroscientists 7. He also discusses the integration of spectral learning with machine learning and physics, emphasizing the elegance of optimization on matrix manifolds 1. Pfau appreciates the mathematical properties of these methods, stating, "I like connecting all these ideas together" 7.