Machine learningML & AI Methods

Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms

The research proposes a new set of topology-based complexity notions that correlate with the generalization gap in deep neural networks, offering a computationally efficient way to predict generalization without test data, and surpassing existing topological bounds across various datasets and models.