Multimodal Datasets with Controllable Mutual Information

The paper presents a self-training framework for detecting offensive content on social media, addressing the challenge of limited labeled data. It utilizes Multi-Agent Vision-Language Models (MA-VLMs) to collaboratively assign pseudo-labels to unlabeled data. The framework distinguishes between Agreed-Unknown and Disagreed-Unknown sets based on classifier and MA-VLM agreement. By simulating dual perspectives, it enhances label reliability and employs a Positive-Negative-Unlabeled (PNU) loss for optimization, aiming to improve detection accuracy in low-resource settings.

The paper discusses the challenges of test-time prompt tuning for vision-language models, highlighting the issue of prompt optimization bias that can lead to suboptimal performance in downstream tasks. It identifies two main causes: the model's focus on entropy minimization, which may overlook prediction accuracy, and data misalignment between visual and textual modalities. To address these issues, the authors propose a new method called Doubly Debiased Test-Time Prompt Tuning, aimed at improving model performance in zero-shot settings.

The paper titled 'Laplace Learning in Wasserstein Space' explores the manifold hypothesis, which suggests that high-dimensional data typically exists within low-dimensional subspaces. The authors investigate graph-based semi-supervised learning methods, particularly focusing on Laplace Learning in the Wasserstein space. This research extends classical graph-based learning algorithms from finite-dimensional Euclidean spaces to infinite-dimensional contexts. The study proves the variational convergence of discrete graph p-Dirichlet energy to its continuum counterpart and characterizes the Laplac…

The study investigates the ability of transformer models to predict long steps in the Collatz sequence, a complex arithmetic function that maps odd integers to their successors. The accuracy of the models varies significantly depending on the base used for encoding, achieving up to 99.7% accuracy for bases 24 and 32, while dropping to 37% and 25% for bases 11 and 3. Despite these variations, all models exhibit a common learning pattern, accurately predicting inputs with similar residuals modulo 2^p.

The article presents a new framework called FairReweighing, which utilizes density estimation to enhance fairness in regression tasks. While AI applications in various sectors have raised concerns regarding transparency and fairness across different demographic groups, most existing research has focused on binary classification. This study introduces a mutual information-based metric to evaluate separation violations and proposes a pre-processing algorithm to ensure fairness in regression models, addressing a relatively underexplored area in AI fairness.

The paper introduces Partial Information Decomposition of Features (PIDF), a novel approach for data interpretability and feature selection. Unlike traditional methods that assign a single importance value to features, PIDF utilizes three metrics: mutual information with the target variable, contribution to synergistic information, and redundant information. The authors evaluate PIDF using both synthetic and real-world data, demonstrating its effectiveness through case studies in genetics and neuroscience.

The article presents a novel approach to volumetric medical image segmentation called Coordinative Ordinal-Relational Anatomical Learning (CORAL). This method addresses challenges in segmenting medical images by utilizing continuous anatomical similarity instead of relying on hard binary thresholds. CORAL aims to improve the accuracy of segmentation by capturing both local and global anatomical structures, thereby enhancing the understanding of anatomical progression across different patients.

Higher-order Neural Additive Models (HONAMs) have been introduced as an advancement over Neural Additive Models (NAMs), which are known for their predictive performance and interpretability. HONAMs address the limitation of NAMs by effectively capturing feature interactions of arbitrary orders, enhancing predictive accuracy while maintaining interpretability, crucial for high-stakes applications. The source code for HONAM is publicly available on GitHub.