The field of artificial intelligence is witnessing significant developments in probabilistic modeling and uncertainty quantification. Researchers are exploring new approaches to capture uncertainty in language models, such as fine-grained conditional probability estimation and probabilistic interactive 3D segmentation. These innovations have the potential to improve the accuracy and reliability of AI systems in various applications, including natural language processing and computer vision. Notably, the introduction of novel frameworks like NPISeg3D and SARTM has demonstrated superior performance in segmentation tasks. Furthermore, the development of methods like Mix-QSAM and UncertainSAM has enabled efficient uncertainty quantification in foundation models like SAM. Overall, these advancements are pushing the boundaries of AI research and paving the way for more robust and trustworthy systems. Noteworthy papers include: Always Tell Me The Odds, which presents a state-of-the-art model for fine-grained probability estimation, and UncertainSAM, which introduces a lightweight post-hoc uncertainty quantification method for the Segment Anything Model.
Advances in Probabilistic Modeling and Uncertainty Quantification
Sources
A Factorized Probabilistic Model of the Semantics of Vague Temporal Adverbials Relative to Different Event Types
Always Tell Me The Odds: Fine-grained Conditional Probability Estimation
Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes
Segment Any RGB-Thermal Model with Language-aided Distillation
What do Language Model Probabilities Represent? From Distribution Estimation to Response Prediction
Ensemble Kalman filter for uncertainty in human language comprehension
am-ELO: A Stable Framework for Arena-based LLM Evaluation
Focus on the Likely: Test-time Instance-based Uncertainty Removal
Polynomial-Time Relational Probabilistic Inference in Open Universes
Mix-QSAM: Mixed-Precision Quantization of the Segment Anything Model
UncertainSAM: Fast and Efficient Uncertainty Quantification of the Segment Anything Model
ULFine: Unbiased Lightweight Fine-tuning for Foundation-Model-Assisted Long-Tailed Semi-Supervised Learning