Pixel-level feature attributions are an important tool in eXplainable AI for Computer Vision (XCV), providing visual insights into how image features influence model predictions. The Owen formula for hierarchical Shapley values has been widely used to interpret machine learning (ML) models and their learned representations. However, existing hierarchical Shapley approaches do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with the actual morphological features. Moreover, no prior Shapley method has leveraged data-aware hierarchies for Computer Vision tasks, leaving a gap in model interpretability of structured visual data. To address this, this paper introduces ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula. ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure tailored for images, the Binary Partition Tree (BPT). By using this data-aware hierarchical partitioning, ShapBPT ensures that feature attributions align with intrinsic image morphology, effectively prioritizing relevant regions while reducing computational overhead. This advancement connects hierarchical Shapley methods with image data, providing a more efficient and semantically meaningful approach to visual interpretability. Experimental results confirm ShapBPT’s effectiveness, demonstrating superior alignment with image structures and improved efficiency over existing XCV methods, and a 20-subject user study confirming that ShapBPT explanations are preferred by humans.

Abstract Despite growing interest in monitoring cognitive states, current studies inadequately address individual differences in physiological reactions. Whereas prior works require extensive data from each individual to personalize the model, the current article explores personalization approaches operating with minimal baseline data. We propose three novel methods to personalize the model with only baseline data available for personalization. Further, we systematically compare those to an existing baseline calibration method, a non-personalized model, and a model using all available data for personalization. We conduct experiments with four open datasets with a total of 170 participants, classifying the cognitive states with a prevalent feature-based approach and a recent large time-series foundation model, MOMENT. The experiments target stress and cognitive load detection in realistic classification tasks, which require models to adapt to a new person. The best classification scores after personalizing with minimal data were around 0.7−0.9 and 0.7 balanced accuracy in binary and three-class tasks, respectively. Two of the proposed personalization methods outperformed the non-personalized model in most cases with the feature-based approach, especially in classification tasks with more than two classes, although their performance remained lower than that of the model using all data for personalization. MOMENT showed little benefit from personalization and performed comparably to the feature-based approach even with a non-personalized model. The findings provide a critical overview of the generalizability and necessity of model personalization with little data, and valuable insights into the development of personalized cognition-aware applications.

We investigate the use of a stratified sampling approach for LIME Image, a popular model-agnostic explainable AI method for computer vision tasks, in order to reduce the artifacts generated by typical Monte Carlo sampling. Such artifacts are due to the undersampling of the dependent variable in the synthetic neighborhood around the image being explained, which may result in inadequate explanations due to the impossibility of fitting a linear regressor on the sampled data. We then highlight a connection with the Shapley theory, where similar arguments about undersampling and sample relevance were suggested in the past. We derive all the formulas and adjustment factors required for an unbiased stratified sampling estimator. Experiments show the efficacy of the proposed approach.

Several explainable AI methods allow a Machine Learning user to get insights on the classification process of a black-box model in the form of local linear explanations. With such information, the user can judge which features are locally relevant for the classification outcome, and get an understanding of how the model reasons. Standard supervised learning processes are purely driven by the original features and target labels, without any feedback loop informed by the local relevance of the features identified by the post-hoc explanations.In this paper, we exploit this newly obtained information to design a feature engineering phase, where we combine explanations with feature values. To do so, we develop two different strategies, named Iterative Dataset Weighting and Targeted Replacement Values, which generate streamlined models that better mimic the explanation process presented to the user. We show how these streamlined models compare to the original black-box classifiers, in terms of accuracy and compactness of the newly produced explanations.

The immune system (IS) represents a complex network of cells and molecules devoted to the protection of individuals from external pathogens, and in terms of complexity, it is only second to the central nervous system. As our knowledge of the IS mechanisms has become more exhaustive, interest has grown in applying modeling and simulation techniques in this context. In particular, among these techniques, the Agent Based Models (ABMs) have been increasingly applied for the IS simulation. One of the major drawbacks of ABMs is represented by the lack of well-defined semantics, which may lead to inconsistent results in comparison to other stochastic approaches. In this paper, we make use of the well-defined semantics and the simulation algorithm for ABMs that we proposed in [1] to implement a few models of the Cancer-Immune System. Comparing ABMs and Gillespie’s Stochastic Simulation Algorithm results we show that our methodology brings coherence among the results of ABMs and SSA.

The Model Checking Contest (MCC) is an annual competition of software tools for model checking. Tools must process an increasing benchmark gathered from the whole community and may participate in various examinations: state space generation, computation of global properties, computation of some upper bounds in the model, evaluation of reachability formulas, evaluation of CTL formulas, and evaluation of LTL formulas. For each examination and each model instance, participating tools are provided with up to 3600 s and 16 gigabyte of memory. Then, tool answers are analyzed and confronted to the results produced by other competing tools to detect diverging answers (which are quite rare at this stage of the competition, and lead to penalties). For each examination, golden, silver, and bronze medals are attributed to the three best tools. CPU usage and memory consumption are reported, which is also valuable information for tool developers.

Finding good variable orders for decision diagrams is essential for their effective use. We consider Multiway Decision Diagrams (MDDs) encoding a set of fixed-size vectors satisfying a set of linear invariants. Two critical applications of this problem are encoding the state space of a discrete-event discrete state system (DEDS) and encoding all solutions to a set of integer constraints. After studying the relations between the MDD structure and the constraints imposed by the linear invariants, we define i $$_\mathrm {Rank}$$ , a new variable order metric that exploits the knowledge embedded in these invariants. We evaluate i $$_\mathrm {Rank}$$ against other previously proposed metrics on a benchmark of 40 different DEDS and show that it is a better predictor of the MDD size and it is better at driving heuristics for the generation of good variable orders.

Summary: Short reads sequencing technology has been used for more than a decade now. However, the analysis of RNAseq and ChIPseq data is still computational demanding and the simple access to raw data does not guarantee results reproducibility between laboratories. To address these two aspects, we developed SeqBox, a cheap, efficient and reproducible RNAseq/ChIPseq hardware/software solution based on NUC6I7KYK mini-PC (an Intel consumer game computer with a fast processor and a high performance SSD disk), and Docker container platform. In SeqBox the analysis of RNAseq and ChIPseq data is supported by a friendly GUI. This allows access to fast and reproducible analysis also to scientists with/without scripting experience. Availability and implementation: Docker container images, docker4seq package and the GUI are available at http://www.bioinformatica.unito.it/reproducibile.bioinformatics.html. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Abstract: This position paper introduces the concept of “artificial co-pilot ” (that is, a driver model), with a focus on driver’s oriented cognitive cars, in order to illustrate a new approach for future intelligent vehicles, which overcomes the limitations of nowadays models. The core consists in adopting the human cognitive frame-work for vehicles, following an artificial intelligent approach to take decisions. This paper illustrates in de-tails these concepts, as they are under development in the EU co-funded project HOLIDES. 1.

The Hybrid Automata Stochastic Logic (HASL) has been recently defined as a flexible way to express classical performance measures as well as more complex, path-based ones (generically called "HASL formulas"). The considered paths are executions of Generalized Stochastic Petri Nets (GSPN), which are an extension of the basic Petri net formalism to define discrete event stochastic processes. The computation of the HASL formulas for a GSPN model is demanded to the COSMOS tool, that applies simulation techniques to the formula computation. Stochastic Symmetric Nets (SSN) are a high level Petri net formalism, of the colored type, in which tokens can have an identity, and it is well known that colored Petri nets allow one to describe systems in a more compact and parametric form than basic (uncolored) Petri nets. In this paper we propose to extend HASL and COSMOS to support colors, so that performance formulas for SSN can be easily defined and evaluated. This requires a new definition of the logic, to ensure that colors are taken into account in a correct and useful manner, and a significant extension of the COSMOS tool.

CSLTA is a stochastic temporal logic for continuous-time Markov chains (CTMC), that can verify the probability of following paths specified by a Deterministic Timed Automaton (DTA). A DTA expresses both logic and time constraints over a CTMC path, yielding to a very flexible way of describing performance and dependability properties. This paper explores a model checking algorithm for CSLTAbased on the translation into a Deterministic and Stochastic Petri Net (DSPN). The algorithm has been implemented in a simple Model Checker prototype, that relies on existing DSPN solvers to do the actual numerical computations.