Search papers, labs, and topics across Lattice.
University of Bonn, Germany Lamarr Institute for Machine Learning and Artificial Intelligence, Germany 22email:
5
0
5
Probabilistic embeddings can boost action segmentation performance by over 20%, overcoming the local optimum pitfalls of deterministic methods.
Ignoring modality-specific signals during inference can lead to significant performance drops in multi-modal OOD detection, but this new hybrid approach turns that insight into a robust solution.
Integrating genomic data with ecological constraints allows for unprecedented accuracy in predicting microbial dynamics and carbon cycling in soil systems.
Generated human motions still look janky? This sphere-based self-intersection loss slashes body-part collisions by nearly 50% while actually *improving* overall motion quality.
Injecting structured visual perturbations into multimodal RL significantly stabilizes training and boosts video reasoning performance, even when answers are superficially similar.
