Abstract
The lack of interpretability of Deep Learning models hinders their deployment in clinical contexts. Case-based explanations can be used to justify these models' decisions and improve their trustworthiness. However, providing medical cases as explanations may threaten the privacy of patients. We propose a generative adversarial network to disentangle identity and medical features from images. Using this network, we can alter the identity of an image to anonymize it while preserving relevant explanatory features. As a proof of concept, we apply the proposed model to biometric and medical datasets, demonstrating its capacity to anonymize medical images while preserving explanatory evidence and a reasonable level of intelligibility. Finally, we demonstrate that the model is inherently capable of generating counterfactual explanations.

Abstract

Abstract
In this paper we study the heart sound segmentation problem using Deep Neural Networks. The impact of available electrocardiogram (ECG) signals in addition to phonocardiogram (PCG) signals is evaluated. To incorporate ECG, two different models considered, which are built upon a 1D U-net - an early fusion one that fuses ECG in an early processing stage, and a late fusion one that averages the probabilities obtained by two networks applied independently on PCG and ECG data. Results show that, in contrast with traditional uses of ECG for PCG gating, early fusion of PCG and ECG information can provide more robust heart sound segmentation. As a proof of concept, we use the publicly available PhysioNet dataset. Validation results provide, on average, a sensitivity of 97.2%, 94.5%, and 95.6% and a Positive Predictive Value of 97.5%, 96.2%, and 96.1% for Early-fusion, Late-fusion, and unimodal (PCG only) models, respectively, showing the advantages of combining both signals at early stages to segment heart sounds.

Abstract
This paper aims at researching the design of a current controller for an interleaved Buck converter used to feed a high current 5 kW Polymer electrolyte membrane (PEM) electrolyzer representing a module stack level. The main challenge is to design a robust controller that ensures operation over a wide range of electrolyzer operating points while guaranteeing control requirements and current sharing between the converters. The developed control scheme ensures responsiveness to the requirements of the grid's ancillary services and control over the dynamics of the electrolyzer. MATLAB/Simulink simulation results with dSPACE compatible models are presented to validate the lead-lag controller, designed using root locus, achieving a ripple current of 0.1 A, a 0.3% steady-state error, and a settling time of 50 ms for a step response.

Abstract
The expectations of Machine Learning systems are becoming increasingly demanding, namely in what concerns the diversity of applications, the expected accuracy, and the pressure for results. However, there are cases in which Human experts are needed to label the data, which may have a significant cost in terms of human resources and time. In these cases, it is often best to learn on-the-fly, without expecting for the whole data to be labeled. Often, it is desirable to guide the Human annotators into focusing on the more relevant instances: this constitutes the so-called active learning. In this paper we propose an approach in which a clustering algorithm is used to find groups of similar instances. Then, the procedure is guided with the objective of favoring the annotation of the groups that are under-represented in the labeled dataset. Results show that this approach leads to models that are, over time, more accurate and reliable.

Abstract
Malicious actors can cause severe damage by exploiting software vulnerabilities. In industrial settings, where critical infrastructures rely on software, handling these vulnerabilities with utmost care is crucial to prevent catastrophic consequences. For this purpose, a cybersecurity awareness platform called Sifu was created. This platform automatically assesses challenges to verify its compliance to secure coding guidelines. Using an artificial intelligence method, an interactive component provides players with solution-guiding hints. This paper presents an improved version of the Sifu platform, which evolves the tool in the following aspects: architecture, data model and user interface. The new platform separates the server and client-side using a REST API architecture. It also accommodates an intrinsic and richer layer of gamification, which explores the concept of game rooms at an organizational and gamification level. Finally, it offers an improved interactive training experience for individuals and organizations through a responsive and intuitive single-page web application.