Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Department of Computing Sciences, Bocconi University, Milano, Italy, Bocconi Institute for Data Science and Analytics (BIDSA)
2
0
2
Thermal noise transforms the optimization landscape, revealing algorithmically accessible regions that enhance generalization in non-convex neural networks.
Forget full-batch gradients: this new pseudo-Langevin dynamics method lets you efficiently sample the parameter space of large neural networks using minibatches, scaling to millions of parameters.