Abstract Cross-modal retrieval requires building a common latent space that captures and correlates information from different data modalities, usually images and texts. Cross-modal training based on the triplet loss with hard negative mining is a state-of-the-art technique to address this problem. This paper shows that such approach is not always effective in handling intra-modal similarities. Specifically, we found that this method can lead to inconsistent similarity orderings in the latent space, where intra-modal pairs with unknown ground-truth similarity are ranked higher than cross-modal pairs representing the same concept. To address this problem, we propose two novel loss functions that leverage intra-modal similarity constraints available in a training triplet but not used by the original formulation. Additionally, this paper explores the application of this framework to unsupervised image retrieval problems, where cross-modal training can provide the supervisory signals that are otherwise missing in the absence of category labels. Up to our knowledge, we are the first to evaluate cross-modal training for intra-modal retrieval without labels. We present comprehensive experiments on MS-COCO and Flickr30k, demonstrating the advantages and limitations of the proposed methods in cross-modal and intra-modal retrieval tasks in terms of performance and novelty measures. We also conduct a case study on the ROCO dataset to assess the performance of our method on medical images and present an ablation study on one of our approaches to understanding the impact of the different components of the proposed loss function. Our code is publicly available on GitHub https://github.com/MariodotR/FullHN.git .

Even after early repair of aortic coarctation (AoCo), life expectancy is reduced due to complications such as hypertension. Invasive diagnostic catheterization is used to evaluate peak-to-peak pressure gradients (PGpp) across the CoAo. Clinically significant PGppare those greater than 20 mmHg under resting conditions, in which case the patient is referred for a second intervention to repair the CoAo. In this study, we demonstrate the feasibility of using Physics-Informed Neural Networks (PINNs) to predict PGppin patients with AoCo non-invasively, based on images obtained from cardiac magnetic resonance imaging. We analyzed a group of 3 patients with CoAo under resting and pharmacological stress conditions. We were able to obtain PGppvalues very close to the actual values obtained by diagnostic catheterization, with an absolute error and average percentage error of 0.57 mmHg and 8.29% for the resting condition, and 4.13 mmHg and 8.63% for the pharmacological stress condition. Our method also successfully identified the only patient who presented a clinically significant PGppunder resting conditions, with differences of less than 1 mmHg.

This paper presents an ingestion procedure into an interoperable repository called ALPACS (Access to Local Picture Archiving and Communication Systems). ALPACS serves clinical and hospital users who can access the repository data through an Artificial Intelligence (AI) application called PROXIMITY 1.0. This paper shows the automated procedure for data ingestion from the medical imaging provider into the ALPACS repository. The data ingestion procedure was successfully applied from the data provider (Hospital Clínico de la Universidad de Chile, HCUCH) by applying a pseudo-anonymization algorithm at the source and respecting the privacy of sensitive patient data. The transfer is done using international communication standards for health systems, allowing the replication of the procedure for other medical imaging provider institutions.

Generative adversarial training has recently raised significant interest in recommender systems. Adversarial pairwise learning, in particular, has led to methods to select and create unobserved training samples that generalize user preferences, increasing the accuracy and robustness of collaborative filtering (CF) models. Despite this success, only some authors have analyzed adversarial sampling’s ability to recommend long-tail and cold-start items, as well as their ability to promote novelty and diversity. These concerns are crucial for modern recommender systems.This paper investigates adversarial pairwise learning in data sparsity scenarios in which most items are consumed by only a few users (long-tail items), and there is a substantial proportion of items without interactions (cold-start items). We found that adversarial sampling increases the bias of CF models toward popular items, resulting in under-recommendation of relevant but less popular long-tail items, poor cold start performance, and low aggregate diversity, i.e., unfair coverage of items among recommendation lists. To address these problems, we propose a multi-modal extension of adversarial pairwise learning that incorporates text and visual information about the items in addition to user-item interaction data. As in the original model, our approach relies on a minimax game. A generative model proposes items for a user considering her visual and textual preferences. Then, a collaborative critic discriminates the suggested items from those already consumed by the user.We conduct experiments on three challenging datasets from the online retail domain in which more than 99.99% of the user-item interactions are unknown, and around 2/3 of the items have less than 5 interactions. We evaluate the advantages of our approach to adversarial and non-adversarial methods, achieving state of the art results in the most complex scenario: the recommendation of new items. Furthermore, we found that the proposed adversarial framework successfully leverages content to make more diverse and novel recommendations.Our code is publicly available on GitHub https://anonymous.4open.science/r/M-APL-D7B7/.

Process Mining (PM) focuses on monitoring and optimizing long-running business processes by examining their execution event logs (usually complex and heterogeneous) to obtain insights and enable data-driven decisions. Several Machine Learning (ML) techniques have been recently proposed to exploit these logs as learning datasets and enable examination of past events and predict future ones, but their black-box nature makes hard for human analysts to interpret their results and recognize the key parts of input data. Attention mechanisms (AM) is an ML technique that does address these shortcomings, but it has been little used for PM. This article describes the design, results and findings of a systematic literature review of attention mechanisms for PM. We addressed three research questions: (a) for which applications are AM used? (b) which kinds of AM are used? and (c) how are AM combined with other ML techniques? An initial search yield 73 papers, and inclusion/exclusion criteria left sixteen, published between 2017 and 2023. Key finding are that: (1) the most common application is sequence prediction, (2) most studies combine global and item-wise attention, added as layers after an encoder generates the continuous representation, and (3) emerging research topics include anomaly detection and data representation. This study shows that attention mechanisms can help process analysts to get some sense of interpretability, and showcases the bright potential for process mining of attention mechanisms, which paradoxically have received little attention themselves.

Predicting the emergence of links in large evolving networks is a difficult task with many practical applications. Recently, the Science4cast competition has illustrated this challenge presenting a network of 64.000 AI concepts and asking the participants to predict which topics are going to be researched together in the future. In this paper, we present a solution to this problem based on a new family of deep learning approaches, namely Graph Neural Networks.The results of the challenge show that our solution is competitive even if we had to impose severe restrictions to obtain a computationally efficient and parsimonious model: ignoring the intrinsic dynamics of the graph and using only a small subset of the nodes surrounding a target link. Preliminary experiments presented in this paper suggest the model is learning two related, but different patterns: the absorption of a node by a sub-graph and union of more dense sub-graphs. The model seems to excel at recognizing the first type of pattern.

Frank-Wolfe algorithms have recently regained the attention of the Machine Learning community. Their solid theoretical properties and sparsity guarantees make them a suitable choice for a wide range of problems in this field. In addition, several variants of the basic procedure exist that improve its theoretical properties and practical performance. In this paper, we investigate the application of some of these techniques to Machine Learning, focusing in particular on a Parallel Tangent (PARTAN) variant of the FW algorithm for SVM classification, which has not been previously suggested or studied for this type of problem. We provide experiments both in a standard setting and using a stochastic speed-up technique, showing that the considered algorithms obtain promising results on several medium and large-scale benchmark datasets.

Training a support vector machine (SVM) requires the solution of a quadratic programming problem (QP) whose computational complexity becomes prohibitively expensive for large scale datasets. Traditional optimization methods cannot be directly applied in these cases, mainly due to memory restrictions. By adopting a slightly different objective function and under mild conditions on the kernel used within the model, efficient algorithms to train SVMs have been devised under the name of core vector machines (CVMs). This framework exploits the equivalence of the resulting learning problem with the task of building a minimal enclosing ball (MEB) problem in a feature space, where data is implicitly embedded by a kernel function. In this paper, we improve on the CVM approach by proposing two novel methods to build SVMs based on the Frank–Wolfe algorithm, recently revisited as a fast method to approximate the solution of a MEB problem. In contrast to CVMs, our algorithms do not require to compute the solutions of a sequence of increasingly complex QPs and are defined by using only analytic optimization steps. Experiments on a large collection of datasets show that our methods scale better than CVMs in most cases, sometimes at the price of a slightly lower accuracy. As CVMs, the proposed methods can be easily extended to machine learning problems other than binary classification. However, effective classifiers are also obtained using kernels which do not satisfy the condition required by CVMs, and thus our methods can be used for a wider set of problems.

We explore a technique to learn Support Vector Models (SVMs) when training data is partitioned among several data sources. The basic idea is to consider SVMs which can be reduced to Minimal Enclosing Ball (MEB) problems in an feature space. Computation of such SVMs can be efficiently achieved by finding a core-set for the image of the data in the feature space. Our main result is that the union of local core-sets provides a close approximation to a global core-set from which the SVM can be recovered. The method requires hence a single pass through each source of data in order to compute local core-sets and then to recover the SVM from its union. Extensive simulations in small and large datasets are presented in order to evaluate its classification accuracy, transmission efficiency and global complexity, comparing its results with a widely used single-pass heuristic to learn standard SVMs.

In this paper, we study a single objective extension of support vector machines for multicategory classification. Extending the dual formulation of binary SVMs, the algorithm looks for minimizing the sum of all the pairwise distances among a set of prototypes, each one constrained to one of the convex-hulls enclosing a class of examples. The final discriminant system is built looking for an appropriate reference point in the feature space. The obtained method preserves the form and complexity of the binary case, optimizing just one convex objective function with m variables and 2m+K constraints, where m is the number of examples and K the number of classes. Non-linear extension are straightforward using kernels while soft margin versions can be obtained by using reduced convex hulls. Experimental results in well-known UCI benchmarks are presented, comparing the accuracy and efficiency of the proposed approach with other state-of-the-art methods.