Roberto Prevete

Neural network training is inherently stochastic, with factors such as weight initialization leading to distinct models despite comparable predictive performance. This phenomenon is commonly associated with the Rashomon effect, which describes the existence of multiple near-optimal models for the same task. Although the Rashomon effect has received increasing attention, it remains unclear whether different sources of training stochasticity contribute similarly or differently to its manifestations. In this work, we present an empirical study of the Rashomon phenomenon along three complementary dimensions: solution-space multiplicity, predictive multiplicity, and decision-basis multiplicity. These dimensions are quantified through the empirical size of the empirical Rashomon set, predictive ambiguity, and agreement between XAI attribution maps, respectively. By independently controlling three standard sources of stochasticity, that are weight initialization, mini-batch data ordering, and dropout, we isolate their contribution to each dimension of the Rashomon phenomenon. Experiments on tabular and image classification benchmarks reveal that these sources affect the three dimensions in different ways. In particular, larger empirical Rashomon sets do not necessarily correspond to greater predictive disagreement or lower explanation agreement, indicating that solution-space, predictive, and decision-basis multiplicity capture complementary rather than interchangeable aspects of the Rashomon effect. Overall, our results show that training stochasticity influences not only predictive performance but also the stability of predictions and explanations, highlighting the importance of identifying the specific sources of stochasticity responsible for different manifestations of the Rashomon phenomenon when assessing the reliability, reproducibility, and interpretability of neural network models.

The eXplainable Artificial Intelligence (XAI) research predominantly concentrates to provide explainations about AI model decisions, especially Deep Learning (DL) models. However, there is a growing interest in using XAI techniques to automatically improve the performance of the AI systems themselves. This paper proposes IMPACTX, a novel approach that leverages XAI as a fully automated attention mechanism, without requiring external knowledge or human feedback. Experimental results show that IMPACTX has improved performance respect to the standalone ML model by integrating an attention mechanism based an XAI method outputs during the model training. Furthermore, IMPACTX directly provides proper feature attribution maps for the model's decisions, without relying on external XAI methods during the inference process. Our proposal is evaluated using three widely recognized DL models (EfficientNet-B2, MobileNet, and LeNet-5) along with three standard image datasets: CIFAR-10, CIFAR-100, and STL-10. The results show that IMPACTX consistently improves the performance of all the inspected DL models across all evaluated datasets, and it directly provides appropriate explanations for its responses.

Machine Learning (ML) has revolutionized various domains, offering predictive capabilities in several areas. However, there is growing evidence in the literature that ML approaches are not always used appropriately, leading to incorrect and sometimes overly optimistic results. One reason for this inappropriate use of ML may be the increasing availability of machine learning tools, leading to what we call the “push the button” approach. While this approach provides convenience, it raises concerns about the reliability of outcomes, leading to challenges such as incorrect performance evaluation. In particular, this paper addresses a critical issue in ML, known as data leakage, where unintended information contaminates the training data, impacting model performance evaluation. Indeed, crucial steps in ML pipeline can be inadvertently overlooked, leading to optimistic performance estimates that may not hold in real-world scenarios. The discrepancy between evaluated and actual performance on new data is a significant concern. In particular, this paper categorizes data leakage in ML, discussing how certain conditions can propagate through the ML approach workflow. Furthermore, it explores the connection between data leakage and the specific task being addressed, investigates its occurrence in Transfer Learning framework, and compares standard inductive ML with transductive ML paradigms. The conclusion summarizes key findings, emphasizing the importance of addressing data leakage for robust and reliable ML applications considering tasks and generalization goals.

This paper presents an innovative approach leveraging Neuronal Manifold Analysis of EEG data to identify specific time intervals for feature extraction, effectively capturing both class-specific and subject-specific characteristics. Different pipelines were constructed and employed to extract distinctive features within these intervals, specifically for motor imagery (MI) tasks. The methodology was validated using the Graz Competition IV datasets 2A (four-class) and 2B (two-class) motor imagery classification, demonstrating an improvement in classification accuracy that surpasses state-of-the-art algorithms designed for MI tasks. A multi-dimensional feature space, constructed using NMA, was built to detect intervals that capture these critical characteristics, which led to significantly enhanced classification accuracy, especially for individuals with initially poor classification performance. These findings highlight the robustness of this method and its potential to improve classification performance in EEG-based MI-BCI systems.

Type 1 diabetes mellitus (T1DM) is characterized by insulin deficiency and blood sugar control issues. The state-of-the-art solution is the artificial pancreas (AP), which integrates basal insulin delivery and glucose monitoring. However, APs are unable to manage postprandial glucose response (PGR) due to limited knowledge of its determinants, requiring additional information for accurate bolus delivery, such as estimated carbohydrate intake. This study aims to quantify the influence of various meal-related factors on predicting postprandial blood glucose levels (BGLs) at different time intervals (15 min, 60 min, and 120 min) after meals by using deep neural network (DNN) models. The prediction models incorporate preprandial blood glucose values, insulin dosage, and various meal-related nutritional factors such as intake of energy, carbohydrates, proteins, lipids, fatty acids, fibers, glycemic index, and glycemic load as input variables. The impact of input features was assessed by exploiting eXplainable Artificial Intelligence (XAI) methodologies, specifically SHapley Additive exPlanations (SHAP), which provide insights into each feature's contribution to the model predictions. By leveraging XAI methodologies, this study aims to enhance the interpretability and transparency of BGL prediction models and validate clinical literature hypotheses. The findings can aid in the development of decision-support tools for individuals with T1DM, facilitating PGR management and reducing the risks of adverse events. The improved understanding of PGR determinants may lead to advancements in AP technology and improve the overall quality of life for T1DM patients.