Prototype-Based Networks (PBNs) are inherently interpretable architectures that facilitate understanding of model outputs by analyzing the activation of specific neurons—referred to as prototypes—during the forward pass. The learned prototypes serve as transformations of the input space into a latent representation that more effectively encapsulates the main characteristics shared across data samples, thereby enhancing classification performance. Crucially, these prototypes can be decoded and projected back into the original input space, providing direct interpretability of the features learned by the network. While this characteristic marks a meaningful advancement toward the realization of fully interpretable artificial intelligence systems, our findings reveal that prototype representations can be deliberately or inadvertently manipulated without compromising the superficial appearance of explainability. In this study, we conduct a series of empirical investigations that demonstrate this phenomenon, framing it as a structural paradox potentially intrinsic to the architecture or its design, which may represent a significant robustness challenge for explainable AI methodologies.

Genetic variation operators in grammar-guided genetic programming are fundamental to guide the evolutionary process in search and optimization problems. However, they show some limitations, mainly derived from an unbalanced exploration and local-search trade-off. This paper presents an estimation of distribution algorithm for grammar-guided genetic programming to overcome this difficulty and thus increase the performance of the evolutionary algorithm. Our proposal employs an extended dynamic stochastic context-free grammar to encode and calculate the estimation of the distribution of the search space from some promising individuals in the population. Unlike traditional estimation of distribution algorithms, the proposed approach improves exploratory behavior by smoothing the estimated distribution model. Therefore, this algorithm is referred to as SEDA, smoothed estimation of distribution algorithm. Experiments have been conducted to compare overall performance using a typical genetic programming crossover operator, an incremental estimation of distribution algorithm, and the proposed approach after tuning their hyperparameters. These experiments involve challenging problems to test the local search and exploration features of the three evolutionary systems. The results show that grammar-guided genetic programming with SEDA achieves the most accurate solutions with an intermediate convergence speed.

Two antagonistic categories can be identified within Artificial Intelligence models: symbolic and subsymbolic. Neurosymbolic artificial intelligence is the term used to describe a set of approaches whose goal is to combine both symbolic and subsymbolic models under a unified approach. Design methods for neurosymbolic systems have been presented throughout the years, focusing mostly on theoretical aspects, such as how knowledge is represented within the system or which learning paradigm should be used. However, existing design approaches do not consider contextual or practical aspects. This work presents a template-based approach for the design of neurosymbolic systems, focusing specifically on the interaction between knowledge-based systems (symbolic) and deep learning models (subsymbolic). The proposed method extends previously addressed aspects (limitations and benefits) with contextual and practical aspects (restrictions and considerations). From this general four-dimensional descriptive template, a specific template is provided describing the parameters of each potential integration between knowledge-based systems and deep learning. Instantiation examples are provided to demonstrate the versatility and applicability of the proposal. In addition, this paper presents an extensive review of the current state-of-the-art in the design of neurosymbolic systems. The benefits of the proposed approach are two-fold. First, it enables high-level description of existing neurosymbolic systems, thus facilitating the search process for research. Secondly, the proposed method simplifies the engineering process and reduces development time because: i) previously existing instances that are already described by templates can be easily reused, or ii) the corresponding integration template can serve as a baseline to develop new systems from scratch.

In the field of artificial intelligence (AI) one of the main challenges today is to make the knowledge acquired when performing a certain task in a given scenario applicable to similar yet different tasks to be performed with a certain degree of precision in other environments. This idea of knowledge portability is of great use in Cyber-Physical Systems (CPS) that face important challenges in terms of reliability and autonomy. This article presents a CPS where unmanned vehicles (drones) are equipped with a reinforcement learning system so they may automatically learn to perform various navigation tasks in environments with physical obstacles. The implemented system is capable of isolating the agents' knowledge and transferring it to other agents that do not have prior knowledge of their environment so they may successfully navigate environments with obstacles. A complete study has been performed to ascertain the degree to which the knowledge obtained by an agent in a scenario may be successfully transferred to other agents in order to perform tasks in other scenarios without prior knowledge of the same, obtaining positive results in terms of the success rate and learning time required to complete the task set in each case. In particular, those two indicators showed better results (higher success rate and lower learning time) with our proposal compared to the baseline in 47 out of the 60 tests conducted (78.3%).

Knowledge Graphs (KGs) are among the most commonly used knowledge representation paradigms, being at the core of tasks such as question answering or recommendation systems. Knowledge Graph Completion (KGC) is one of the key tasks concerning KGs, where the goal is to extract new elements from the existing information. Different approaches have been proposed through the years to tackle this chal-lenge. Among them, two analogous categories can be distinguished: rule-learning and Knowledge Graph Embeddings (KGE). Different methods have been subsequently proposed to unify both types under a sin-gle framework, such that the benefits of both proposals can be exploited. However, most of these meth-ods consider using rule-learning models as a boosting agent for KGE models, but not as an explainability tool. This work presents GEnI1, a framework capable of generating insights and explanations for KGE mod -els. GEnI follows a three-phase sequential process, generating a feasible explanation for a given prediction. Possible outcomes are rules, correlations, and influence detection. Moreover, the output is expressed in nat-ural language to further extend the explainability of the proposal. GEnI has been successfully evaluated under three criteria: coherence, the meaningfulness of the output, and reliability. Moreover, it can be used by both translational and bilinear KGE models, offering broad coverage. Furthermore, this work also pre-sents an in-depth review of existing integrative approaches between rule-learning and embedding models, providing a comparative framework between them. (c) 2022 Elsevier B.V. All rights reserved.

Knowledge graphs (KGs) are one of the most common frameworks for knowledge representation. However, they suffer from a severe scalability problem that hinders their usage. KG embedding aims to provide a solution to this issue. Nonetheless, general approaches are incapable of representing and reasoning about information not previously contained in the graph. This paper proposes to leverage semantic and ontological information for a significant benefit of knowledge graph completion, focusing on triple classification. The goal of this task is to determine whether a given fact holds. Furthermore, this paper also considers the classification of facts that include entities that have not been seen during training, denoted out-of-knowledge-base or OOKB entities. An incremental method is presented, composed of six stages. Although the proposal can be applied to any KG embedding model, this work focuses on its application for semantic matching models, such as ComplEx and DistMult. Compared to other approaches, our proposal is model-agnostic, computationally inexpensive, and does not require retraining. The results show that triple classification accuracy scales up to 15% with the proposed approach, as well as accelerating the convergence of the model to its optimal solution. Furthermore, facts containing OOKB entities can be classified with a reasonable accuracy. (c) 2021 Elsevier Inc. All rights reserved.

Case-Based Reasoning models are one of the most used reasoning paradigms in expert-knowledge-driven areas.One of the most prominent fields of use of these systems is the medical sector, where explainable models are required.However, these models are considerably reliant on user input and the introduction of relevant curated data.Deep learning approaches offer an analogous solution, where user input is not required.This paper proposes a hybrid Case-Based Reasoning, Deep Learning framework for medical-related applications, focusing on the generation of medical reports.The proposal combines the explainability and user-focused approach of case-based reasoning models with the deep learning techniques performance.Moreover, the framework is fully modular to fit a wide variety of tasks and data, such as real-time sensor captured data, images, or text, to name a few.An implementation of the proposed framework focusing on radiology report generation assistance is provided.This implementation is used to evaluate the proposal, showing that it can provide meaningful and accurate corrections, even when the amount of information available is minimal.Additional tests on the optimization degree of the case base are also performed, evidencing how the proposed framework can optimize this base to achieve optimal performance.

Hyper-personalization policies entail a considerable improvement regarding previous personalization approaches. However, they present several issues that need to be addressed, such as minimal explainability and privacy invasion. A hierarchical Multi-Agent System (MAS) is presented in this work to provide a solution to these concerns. The system is formulated as a hybrid approach, where some of the agents work autonomously, while the user input triggers the remaining. At the autonomous level, a set of Virtual Identities (VIs) representing different user profiles interact with Black-Box Hyper-Personalization Online Systems (BBHOS), gathering a set of targeted responses. Associative patterns and profile aggregations can then be inferred from the analysis of these responses. In the user-triggered level, the real user is virtualized as an identity that represents their features. The virtual identity serves as an intermediary between the personalization system and the real user. This virtualization hinders the personalization service from extracting sensitive contextual information about the real user, protecting their privacy. The results obtained by the user identity on its interaction with the personalization service are then analyzed, adjusting the content of the response to fit the user's requests instead of their features. A use case on the functioning of the analysis of search engines is presented to illustrate the complete behavior of the proposed architecture.

For a fixed p is an element of N, sequences of polynomials {P-n}, n is an element of N, defined by a ( p + 2)-term recurrence relation are related to several topics in Approximation Theory. A (p + 2)-banded matrix J determines the coefficients of the recurrence relation of any of such sequences of polynomials. The connection between these polynomials and the concept of orthogonality has already been established through a p-dimension vector of functionals. This work goes further on this topic by analyzing the relation between such vectors for the set of sequences {P-n((j))}, n is an element of N, associated with the Darboux transformations J((j),) j = 1,..., p, of a given ( p + 2)-banded matrix J. This is synthesized in Theorem 1, where, under certain conditions, these relationships are established. Besides, some relationships between the sequences of polynomials {P-n((j))} are determined in Theorem 2, which will be of interest for future research on p-orthogonal polynomials. We also provide an example to illustrate the effect of the Darboux transformations of a Hessenberg banded matrix, showing the sequences of p-orthogonal polynomials and the corresponding vectors of functionals. For the sake of clarity, in this example we have considered the case p = 2, since the procedure is similar for p > 2. (C) 2020 Elsevier Inc. All rights reserved.

The initial population distribution is an essential issue in evolutionary computation performance. Population initialization methods for grammar-guided genetic programming have some difficulties generating a representative sample of the search space, which negatively affects the overall evolutionary process. This paper presents a grammatically uniform population initialization method to address this issue by improving the initial population uniformity: the equiprobability of obtaining any individual of the search space defined by the context-free grammar. The proposed initialization method assigns and updates probabilities dynamically to the production rules of the grammar to pursue uniformity and includes a code bloat control mechanism. We have conducted empirical experiments to compare the proposed algorithm with a standard initialization approach very often used in grammar-guided genetic programming. The results report that the proposed initialization method approximates very well a uniform distribution of the individuals in the search space. Moreover, the overall evolutionary process that takes place after the population initialization performs better in terms of convergence speed and quality of the final solutions achieved when the proposed method generates the initial population than when the usual approach does. The results also show that these performance differences are more significant when the experiments involve large search spaces.

A moderate experiential learning is proposed as a framework to provide learners with significant experiences in data science. In this approach, the student learns through reflection on doing, abstract conceptualization, gamification and learning transferring; instead of being a recipient of already made content. Data science pedagogy has repeated a number of patterns that can be detrimental to the student. The proposed moderate experiential learning has been adopted together with other two learning approaches in a data science master subject for comparative purposes: a traditional learning approach, and a strict experiential learning adoption. Two evaluation studies have been conducted to compare these three different learning approaches. The results indicate that students do not actively support the strict experiential learning, but the moderate approach, where some guidelines are provided to face the realistic experience.

This paper introduces a new initialization method for knowledge graph (KG) embedding that can leverage ontological information in knowledge graph completion problems, such as link classification and link prediction. Although the initialization method is general and applicable to different KG embedding approaches in the literature, such as TransE or RESCAL, this paper experiments with deep learning and specifically with the neural tensor network (NTN) model. The experimental results show that the proposed method can improve link classification for a given relation by up to 15%. In a second contribution, the proposed method allows for addressing a problem not studied in the literature and introduced here as “KG completion with fresh entities”. This is the use of KG embeddings for KG completion when one or several of the entities in a triple (head, relation, tail) has not been observed in the training phase.

Ambient Intelligence is currently a lively application domain of Artificial Intelligence and has become the central subject of multiple initiatives worldwide. Several approaches inside this domain make use of knowledge bases or knowledge graphs, both previously existing and ad hoc. This form of representation allows heterogeneous data gathered from diverse sources to be contextualized and combined to create relevant information for intelligent systems, usually following higher level constraints defined by an ontology. In this work, we conduct a systematic review of the existing usages of knowledge bases in intelligent environments, as well as an in-depth study of the predictive and decision-making models employed. Finally, we present a use case for smart homes and illustrate the use and advantages of Knowledge Graph Embeddings in this context.

Data Science is an interdisciplinary field devoted to extract knowledge from large amounts of data. There is a great variety of pro- grams that address the teaching of this field with a growing demand of professionals. However, data science pedagogy tends to emphasize gen- eral aspects of data and the use of tools instead of the its scientific dimension. This position paper describes an ongoing educational inno- vation project for the use of the Challenge-based Learning approach to teach and learn Data Science. In this approach, students work on solv- ing complex and real world problems while the learning is obtained by iterating through three main phases: engage, investigate, and act.

This article presents an evolutionary algorithm to autonomously construct full-connected multilayered feedforward neural architectures. This algorithm employs grammar-guided genetic programming with a context-free grammar that has been specifically designed to satisfy three important restrictions. First, the sentences that belong to the language produced by the grammar only encode all valid neural architectures. Second, full-connected feedforward neural architectures of any size can be generated. Third, smaller-sized neural architectures are favored to avoid overfitting. The proposed evolutionary neural architectures construction system is applied to compute the terms of the two sequences that define the three-term recurrence relation associated with a sequence of orthogonal polynomials. This application imposes an important constraint: training datasets are always very small. Therefore, an adequate sized neural architecture has to be evolved to achieve satisfactory results, which are presented in terms of accuracy and size of the evolved neural architectures, and convergence speed of the evolutionary process.

La Ciencia de Datos es una revolución que ya está cambiando la manera en la que nos ocupamos de sanidad, política, educación e innovación. Hay una gran variedad de cursos online, másteres, grados, y asignaturas que se enfocan a la enseñanza de este campo interdisciplinar, donde existe una demanda creciente de profesionales. Este artículo describe tres modelos de enseñanza y aprendizaje diferentes para Ciencia de Datos, inspirados en el paradigma del aprendizaje experiencial. Además, estos modelos han sido empleados en una asignatura de Deep Learning, dentro del contexto de un máster internacional de Ciencia de Datos. Finalmente, la satisfacción de los estudiantes es cuantificada por medio de una encuesta para: evaluar los tres modelos, decidir el enfoque favorito y considerar una metodología de aprendizaje experiencial pura.X49 [Data science is a revolution that is already changing the way we do business, healthcare, politics, education and innovation. There is a great variety of online courses, masters, degrees, and modules that address the teaching of this interdisciplinary field, where is a growing demand of professionals. This paper describes three different teaching and learning models for Data Science inspired by the experiential learning paradigm. Moreover, these models have been employed in a Deep Learning course in the context of an international master of data science. Finally, the student satisfaction is quantified by a survey to: evaluate the three models, decide their favorite approach, and consider a pure experiential learning methodology. ] Serrano, Emilio; Molina, Martin; Manrique, Daniel; Baumela, Luis; Zanardini, Damiano

Data science is a revolution that is already changing the way we do business, healthcare, politics, education and innovation. There is a great variety of online courses, masters, degrees, and modules that address the teaching of this interdisciplinary field, where is a growing demand of professionals. However, data science pedagogy has repeated a number of patterns that can be detrimental to the student. This position paper describes an ongoing educational innovation project for the study of methods, experiences, and tools for experiential learning in data sicience. In this approach, the student learns through reflection on doing instead of being a recipient of already made content.

In this paper we explore the hybrid application of evolutionary computation and artificial neural networks in the development of intelligent systems able to solve the problem of approximating the optimal strategy in a tile-matching puzzle game. Three intelligent systems are proposed: an evolutionary heuristic technique, artificial neural networks, and a hybrid approach that combines both. Results show that the hybrid approach, which combines the advantages of the two previous solutions, performs better at both, the number of completed lines and the average piece placement time. These results aim to serve as the basis for a later comparative study against state- of-the-art techniques in the topic.

Encoding feasible solutions is one of the most important aspects to be taken into account in the field of evolutionary computation in order to solve search or optimization problems. This paper proposes a new encoding scheme for real-coded evolutionary algorithms. It is called partition based encoding scheme, and satisfies two restrictions. Firstly, each of the components of a decoded vector that conforms a candidate solution to a problem at hand belongs to a predefined interval. Secondly, the sum of the components of each of these decoded vectors is always equal to a predefined constant. The proposed encoding scheme inherently guarantees these constraints for all the individuals that are generated within the evolution process as a consequence of applying the genetic operators. Partition based encoding scheme is successfully applied to learning conditional probability tables for a given discrete Bayesian network topology, where each row of the tables must exactly add up to one, and the components of each row belong to the interval [0,1] as they are probability values. The results given by the proposed encoding system for this learning problem is compared to a deterministic algorithm and another evolutionary approach. Better results are shown in terms of accuracy with respect to the former one, and accuracy and convergence speed with respect to the later one.