The scarcity of medical data, particularly in Survival Analysis (SA) for cancer-related diseases, challenges data-driven healthcare research. While Synthetic Tabular Data Generation (STDG) models have been proposed to address this issue, most rely on datasets with abundant samples, not reflecting real-world limitations. We suggest using an STDG approach that leverages transfer learning and meta-learning techniques to create an artificial inductive bias, guiding generative models trained on limited samples. Initial experiments were conducted on larger classification datasets, allowing us to asses the methodology across varying sample sizes and abundant versus scarce data scenarios. We primarily employed clinical utility validation for cancer-related SA data, as divergence-based similarity validation was not feasible. The methodology improved STDG under constrained data conditions, with divergence-based similarity validation proving to be a robust measure of data quality. Conversely, clinical utility validation yielded similar results regardless of sample size, indicating its limitations in statistically confirming effective STDG. In SA experiments, we observed that altering the task can reveal if relationships among variables are accurately generated, with most cases benefiting from the proposed methodology. Our study underscores the efficacy of the approach in tackling medical data scarcity by effectively generating high-quality synthetic data under constrained conditions. While divergence-based similarity validation is essential when sufficient data are available, clinical utility validation alone is insufficient and should be complemented by similarity validation. These findings underscore the potential and limitations of STDG methodologies in addressing medical data scarcity.

We introduce DeCaFlow, a deconfounding causal generative model. Training once per dataset using just observational data and the underlying causal graph, DeCaFlow enables accurate causal inference on continuous variables under the presence of hidden confounders. Specifically, we extend previous results on causal estimation under hidden confounding to show that a single instance of DeCaFlow provides correct estimates for all causal queries identifiable with do-calculus, leveraging proxy variables to adjust for the causal effects when do-calculus alone is insufficient. Moreover, we show that counterfactual queries are identifiable as long as their interventional counterparts are identifiable, and thus are also correctly estimated by DeCaFlow. Our empirical results on diverse settings—including the Ecoli70 dataset, with 3 independent hidden confounders, tens of observed variables and hundreds of causal queries—show that DeCaFlow outperforms existing approaches, while demonstrating its out-of-the-box applicability to any given causal graph.

While synthetic tabular data generation using Deep Generative Models (DGMs) offers a compelling solution to data scarcity and privacy concerns, their effectiveness relies on the availability of substantial training data, often lacking in real-world scenarios. To overcome this limitation, we propose a novel methodology that explicitly integrates artificial inductive biases into the generative process to improve data quality in low-data regimes. Our framework leverages transfer learning and meta-learning techniques to construct and inject informative inductive biases into DGMs. We evaluate four approaches (pre-training, model averaging, Model-Agnostic Meta-Learning (MAML), and Domain Randomized Search (DRS)) and analyze their impact on the quality of the generated text. Experimental results show that incorporating inductive bias substantially improves performance, with transfer learning methods outperforming meta-learning, achieving up to 60% gains in Jensen-Shannon divergence. The methodology is model-agnostic and especially relevant in domains such as healthcare and finance, where high-quality synthetic data are essential, and data availability is often limited.

Federated Learning is an emerging, powerful approach that allows training an artificial intelligence model in distributed setting. Two survival models, Cox and DeepSurv, have been trained in a federated setting, exploiting both code simulations and real experiments on the new platform, developed by the GenoMed4All consortium. Different scenarios have been tested by splitting a Myelodysplastic Syndrome dataset into three nodes and performing feature removal. A significant gain in model performance has been observed due to federated aggregation.

Discounted repeated games are currently being used to model the conflicts that arise between the nodes in a wireless network, such as distributed resource allocation, interference management or defending the network against attacks. In current literature, it is frequent that authors devise a specific strategy that performs well only for their concrete problem, thus, it would be desirable to have a generic algorithm that allows learning strategies for such games. However, current learning algorithms focus on average payoff repeated games, and we show analytically that there are important differences that prevent us from using such algorithms for discounted repeated games. In this work, we aim to fill this gap and we propose LEWIS, a lightweight, online learning algorithm specifically designed for these games, that deals with imperfect and incomplete information and is able to return a good payoff. We test LEWIS on two settings based on current literature problems to show that it has a good performance, hence, being a promising method to learn how to play a discounted repeated game in wireless networks.

The recent advances in Deep Learning have a significant impact on the security of wireless networks, such as intelligent attackers which are able to successfully exploit a possibly unknown defense mechanism simply by interacting with it. Their capacity of adapting to standard defense mechanisms, such as statistical tests, makes them a significant threat. In this article, we develop two intelligent defense mechanisms using inverse reinforcement learning tools, that can be used to enhance the capabilities of current defense mechanisms. We test our proposal on a backoff attack setup against an intelligent attacker, obtaining very significant gains in the defense performance.

Safety in coastal areas such as beaches, ports, pontoons, etc., is a current problem with a difficult solution and on which many organizations are putting efforts in terms of technological innovation. In this work the design of a possible solution based on magnetic sensors is presented. First, a study has been made of the type of sensors that best suit the application based on parameters such as sensitivity, the allowed bandwidth of excitation, price or physical construction. Then the system of excitation of the sensors and signal measurement is presented. To justify the design, a series of simulations of magnetic field variations have been carried out in the presence of large objects of conductive material, in the vicinity of the measuring points. With these data a mathematical model has been established that allows the identification of the dimensions and position of the object through triangulation and knowing only the data of the magnetic field. It was found that although this method seems quite effective, it has a significant error, so another method based on neural networks was developed also using data from the simulations. This method seems to yield much better and more reliable results.

We study an underwater navigation problem, where an Under-water Autonomous Vehicle must reach a target position in the presence of a disturbance that may be unknown. In order to deal with this problem, we make use of Deep Reinforcement Learning tools, and more concretely, we make use of robust control ideas, which allow training an agent in the presence of uncertainty. We propose a robust Proximal Policy Optimization agent and train it using simulations of an underwater medium: this agent shows an excellent performance when facing unknown disturbances, being able to approach the performance of the optimal agent which had an exact knowledge of the underwater disturbance.

In this work, we develop an algorithm based on graph networks to train distributedly a deep learning model. We consider that there are several nodes, in an arbitrary network topology, each one of them having access to a local dataset that, for privacy concerns, cannot be shared with other nodes. We also assume that there are bandwidth constraints, and hence, it may not be affordable sharing the weights of the model with other nodes. Our algorithm makes use of pruning and an autoencoder to train under all these constraints, and our simulations show that it provides a good accuracy, while preserving the privacy of the data and providing a compression of nearly two orders of magnitude in the transmitted bits.