Academic papers and research publications. Google Scholar profile →
Efficient search through structured kernel spaces for Gaussian processes using a kernel-kernel over symbolic representations and optimal transport, avoiding costly function-space distances.
TL;DR: A kernel family with learnable input-partitioning for active learning of nonstationary GPs: learnable via gradients, flexible geometry, and applicable in low-data regimes as a good prior for active learning.
Addresses the computational bottleneck of learning kernel parameters by amortizing inference over datasets and kernel structures with a transformer-based network that accepts datasets and symbolic kernel descriptions.
TL;DR: Comparative analysis of closed-loop training for realistic traffic agents: open vs. closed-loop training, generative adversarial vs. deterministic supervised methods, impact of RL losses, and training with log-replayed agents in highway driving simulations.
For a complete list with citations and metrics, see Google Scholar.