NOVEL TECHNIQUES FOR HARNESSING SYMBOLIC AND STRUCTURED INFORMATION INTO MACHINE LEARNING - Dottorato di ricerca in Computer Science and Engineering (Ciclo 36)

Aug 3, 2026

Upload by Ezra

The thesis analyzes how recent advances in deep learning shifted AI from traditional symbolic methods toward sub-symbolic, data-driven approaches, enabled by dedicated hardware such as GPUs/TPUs and by the rapid growth of available data, often coming from IoT sensor streams. It addresses limitations of purely sub-symbolic models, including redundant learning of known concepts, difficulty enforcing explicit constraints, and limited interpretability of black-box systems. It proposes and investigates informed machine learning, combining formalized task knowledge with machine learning workflows to improve efficiency, reliability, adaptability, and explainability through predictive modeling and decision support.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages132

Reading Time5 h 33 min

AccessFree to download

Abstract

Introduction

Background

Optimization under constraints

Optimization under Uncertainty

Machine Learning

Informed Machine Learning

Algebraic equations

Differential equations

Logic rules

Simulation results

Bayesian networks

Declarative formulation of an optimization problem

Use cases

Energy Management System

Predictive Maintenance

Resistor Capacitor circuit for Thermal Modeling

Combinatorial Optimization

Knowledge injection methods to improve predictive models

External model integration

Universal Differential Equation for data-driven discovery of ODEs

DOTTORATO DI RICERCA IN COMPUTER SCIENCE AND ENGINEERING Ciclo 36 Settore Concorsuale: 09/H1-SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI Settore Scientifico Disciplinare: ING-INF/05-SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI NOVEL TECHNIQUES FOR HARNESSING SYMBOLIC AND STRUCTURED INFORMATION INTO MACHINE LEARNING Presentata da: Mattia Silvestri Coordinatore Dottorato Ilaria Bartolini Supervisore Michele Lombardi Esame finale anno 2024 Abstract In recent years, we have assisted to a new spring of Artificial Intelligence (AI) . This transformation has been characterized by a shift from the symbolic methods prevalent in the last century to a focus on sub-symbolic techniques, driven by the remarkable achievements of deep learning in areas such as computer vision and natural language processing. The renewed interested in AI is attributed to two pivotal factors: the advent of powerful, dedicated hardware like graphical processing units and tensor processing units, providing the computational power necessary for training complex deep learning models, and the enormous data availability in the age of big data. The latter, especially augmented by the widespread adoption of Internet of Things (IoT) technologies, has led to an abundance of data from diverse sources, particularly sensor measurements, fueling the application of machine learning and deep learning in various industrial scenarios. Despite the successes of sub-symbolic, data-driven methods, recent years have seen a growing inclination towards hybrid models that synergize symbolic and sub-symbolic approaches. This trend stems from several inherent limitations in purely data-driven systems. Firstly, these systems often redundantly learn concepts that are already part of common knowledge or are well-understood by domain experts. This redundancy raises the question of how to prevent machine learning algorithms from re-learning these established concepts. Secondly, data-driven methods may struggle to adhere to specific constraints, such as those dictated by natural laws or user-imposed rules, whereas symbolic methods can manage these constraints more easily. Lastly, the black-box nature of sub-symbolic methods poses challenges in terms of interpretability and explainability, in contrast to the more transparent symbolic approaches. In the context machine learning and deep learning, these challenges have given rise to the emergent field of informed machine learning. This new domain aims to exploit the strengths of both symbolic and sub-symbolic methods by formalizing and incorporating existing task-specific knowledge into traditional machine learning workflows. The goal is to create systems that are not only more efficient and reliable but also more interpretable and adaptable to various constraints. The core objective of this thesis is to explore and advance the field of informed machine learning. It presents innovative algorithms within this domain and conducts a thorough investigation of existing methodologies. The applications of these algorithms are explored in two significant areas of AI: predictive modeling and decision support systems. To validate the practical utility of these algorithms, the thesis undertakes a comprehensive empirical evaluation. This evaluation encompasses real-world application as well as abstract problems commonly used in the scientific community to investigate practical use cases. The findings from these studies provide concrete evidence of the effectiveness of informed machine learning solutions in addressing the highlighted challenges. Moreover, the thesis demonstrates how informed machine learning can significantly enhance the capabilities and applicability of sub-symbolic methods by effectively harnessing diverse forms of existing knowledge. Table of contents List of figures 7 List of tables 11 Research activities and publications 13 1 Introduction 15 2 Background 17 2. 1 Optimization under constraints 1

FAQs

Which two main application areas does the thesis focus on?

The algorithms are explored in predictive modeling and in decision support systems, validated through comprehensive empirical evaluation on both real-world applications and abstract scientific problems.

Aug 3, 2026

Upload by Ezra