Discovering and disambiguating entities from multimodal and heterogeneous data is a key challenge in discovery science, where relevant knowledge is often implicit and distributed across modalities. We present KARAMEL, a framework for multimodal entity linking (MEL)—the task of aligning visual and textual mentions to entities in a knowledge base—in the cultural heritage domain. Unlike prior vision–language models that primarily rely on surface-level multimodal alignment, our approach incorporates sociohistorical and contextual knowledge through a knowledge-aware architecture that fuses a vision–language encoder with a heterogeneous graph neural network. This design enables contextual reasoning over complex relational structures and supports robust entity disambiguation under ambiguous or underspecified evidence. To benchmark multimodal discovery in this setting, we introduce 𝒲 iki ℳ u 𝒮𝒜 ( 𝒲 ikidata-based ℳ ultimodal 𝒮 emantic data for 𝒜 rt), a new dataset that links artworks to Wikidata entities via images, textual descriptions, and structured contextual knowledge. Experimental results on both 𝒲 iki ℳ u 𝒮𝒜 and MELArt show that KARAMEL surpasses strong baselines, particularly on implicit or abstract mentions, while ablation studies highlight the critical role of domain-specific, structured context in learning expressive representations. These results suggest that KARAMEL has potential for other scientific applications where knowledge is critical for better entity linking. Both the dataset and the code are publicly available on Zenodo ( https://zenodo.org/records/23019880 ) and GitHub ( https://github.com/cilabuniba/karamel ).
Regularization is a central and fundamental research topic in machine learning. Despite the numerous methods developed in current fast growing literature, there is a lack of theoretical understanding and rigorous analysis of regularization. The first part of this paper is an answer to this need: we formulate regularization as a solution of ill-posed problems, examine fundamental principles of regularization, analyze the effectiveness of Tikhonov regularization in linear systems, and propose a consistency principle for the design of regularization methods in both linear and nonlinear systems. Based on these analyses, the second part of this paper is a survey of recent loss-based regularizers. We develop a comprehensive taxonomy and summarize prominent features of these regularizers. Finally, results of benchmark tests confirm the effectiveness of representative loss-based regularizers.
Decision tree learning algorithms, such as Classification and Regression Trees (CART), are commonly founded on heuristics that greedily maximize purity gain. Despite their practical success, the theoretical properties of these algorithms, such as consistency, remain unclear. This paper uncovers a substantial obstacle in analyzing the consistency of decision tree learning algorithms, in which the worst-case purity gain can be zero and a gain value of zero usually results in a non-convergent training error. Building upon this observation, our previous work introduced GridCART, a grid-based decision tree that can yield positive purity gain and achieve consistency under Lipschitz smoothness and a product feature distribution. This article extends the grid-based construction to random forests and gradient boosting decision trees, and establishes consistency-rate upper bounds for Grid-based Random Forests (GridForest) and a square-loss specialization of the Grid-based Gradient Boosting Decision Tree (GridGBDT). Experiments on 20 OpenML datasets show that the grid-based ensembles perform comparably to their conventional counterparts.
Offline Reinforcement Learning provides a viable solution for tasks where online interaction is costly or infeasible by leveraging pre-collected data for policy learning. However, in an offline setting, the policy improvement process is prone to value overestimation due to distribution shifts, which can lead to policy degradation. Existing methods primarily address this issue through policy constraints or conservative value regularization, but few explicitly model the problem actions from a representation learning perspective. This paper proposes TACCO (TD3+BC with Actor-Critic Contrastive Optimization), a contrastive offline reinforcement learning method based on the TD3+BC framework, designed to address the out-of-distribution (OOD) action problem in offline reinforcement learning. We define OOD actions as those that lack local data support in a given state, do not match the current state, or exhibit low reliability. Furthermore, for dense and sparse reward environments, we design two distinct contrastive learning policies: TACCO-A performs contrastive learning on the actor side between data-driven actions, random actions, and cross-state policy actions; TACCO-B filters reliable samples on the critic side based on high Q-values and low uncertainty, and applies contrastive constraints to both reliable and remaining samples simultaneously. In this paper, we compare TACCO with six state-of-the-art offline reinforcement learning algorithms on the Mujoco and Adroit environments within the D4RL benchmark. Experimental results demonstrate that TACCO achieves superior performance on most tasks, validating its effectiveness in mitigating OOD action interference and enhancing policy stability.
Adversarial training improves model robustness but may exhibit robust overfitting. Previous theoretical analyses mostly emphasize Lipschitz continuity of the adversarial loss, while understanding under weaker smoothness conditions (often observed in practice) remains limited. This paper investigates this problem in adversarial training, focusing on the generalization and optimization analyses of adversarial losses under weaker smoothness. We introduce Hölder continuity for the adversarial loss and propose a new concept of approximate Hölder smoothness, extending classical analyses from Lipschitz assumptions to a more general regime. A stability analysis for adversarial training is then developed, forming the theoretical basis for analyzing generalization and optimization. Under these conditions, we establish upper bounds on the generalization error and optimization gap of stochastic gradient descent for both convex and non-convex adversarial losses, and analyze the induced excess risk. These findings apply to a broad class of non-Lipschitz but Hölder smooth losses in deep learning, and offer theoretical insights into how optimization and generalization interact in adversarial training, contributing to a systematic understanding of robust overfitting. Furthermore, we empirically validate our theoretical findings using a practical example, demonstrating that the framework accurately captures generalization behavior under adversarial training.
