Search papers, labs, and topics across Lattice.
Department of Statistical Science, University College London, UK
2
0
3
Neural posteriors are often overconfident with limited data, but DRO-NPE offers a surprisingly simple fix: train against a worst-case loss within a Wasserstein ball.
Get faster convergence in entropy-regularized learning tasks: Kernel Thinning slashes the computational cost of Mean Field Langevin Dynamics while preserving its theoretical guarantees.
