International Summer School on Artificial Intelligence, Data and Research

Sihem Amer-Yahia — AI Planning for Data Exploration

CNRS Research Director and Deputy Director of the Lab of Informatics of Grenoble

Abstract

Exploratory Data Analysis (EDA) is an online decision making process whereas the next best step is chosen based on the latest observed insights. AI Planning is a long standing sub-area of AI, dealing with sequential decision making. This course examines the applicability of two main AI Planning approaches to EDA: Reinforcement Learning (RL) and Agentic AI.

RL4EDA trains an agent to choose the best query, a.k.a. action, on data, and rewards the agent for found insights. The first part of this course will cover the use of various RL algorithms such as SARSA and Actor/Critic, to solve common EDA questions such as finding galaxies of interest or navigating customer reviews.

The second part of the course will introduce Agentic AI, a nascent area where a supervisor agent orchestrates multiple specialized ones to achieve a task. We will introduce a reference architecture for Agentic AI that includes task decomposition and agent discovery.

The parallel with RL immediately comes to mind and that is what we plan to explore in the hands-on part of this course where students will deploy both RL and Agentic AI and get to experience their potential and limitations for finding items of interest.

The last part of the course will dive into multiple open problems that arise when seeking to close the gap between RL4EDA and Agentic AI for EDA.

Biography

Sihem Amer-Yahia is a Silver Medal CNRS Research Director and Deputy Director of the Lab of Informatics of Grenoble. She works on exploratory data analysis and algorithmic upskilling. Prior to that she was Principal Scientist at QCRI, Senior Scientist at Yahoo! Research and Member of Technical Staff at AT&T Labs. Sihem served as PC chair for SIGMOD 2023 and as the coordinator of the Diversity, Equity and Inclusion initiative for the database community. In 2024, she received the 2024 IEEE TCDE Impact Award, the SIGMOD Contributions Award, and the VLDB Women in Database Award.

Walid Aref — Data Indexing in the One- and Multi-Dimensional Spaces

Purdue University

In this session, we will have an overview of the field of database indexing. The focus will be indexing one- and multi-dimensional space. For example, can we index all of Ivory Coast's road network and answer queries on the vehicles moving in these roads?

More specifically, we will cover the following topics:

• Introduction to Data Indexing
• What is the B-tree and Why is the B-tree so Popular?
• Does There Exist a B-tree-like Index for Multi-dimensional Data, e.g., For Indexing Road Maps and for Moving Vehicles?
• Fundamental Approaches to Multi-Dimensional Data Indexing
  • Space-Filling Curve-based Approaches
  • The R-tree and GiST (Generalized Search Trees)
  • SP-GiST (Space-Partitioning Generalized Search Trees)
  • Supporting the Temporal Dimension (e.g., For Indexing Moving Vehicles)

Walid Aref — Databases for AI and AI for Databases: The Story of Two Systems

Purdue University

Databases and AI support each other in two interesting directions:

• Databases for AI: databases can help manage and support data that is being heavily used by ML and LLM models.
• AI for Databases: ML and LLM models can help optimize the performance of database engines and replace many traditional database techniques that involve heuristics with ML and LLM models.

In this session, we will study both directions.

Part I — Databases for AI

We will demonstrate why current databases are at a disadvantage when it comes to using ML and LLM tools. We will address the following challenges:

• How to bring ML and LLM functionalities into the inside of the database management systems?
• How to extend the SQL query language to support in-database ML and LLM inferencing?
• How to use LLM and ML functions to extend the traditional relational algebra operations to support their semantic counterparts, e.g., semantic projects, semantic selects, semantic joins, or for semantic grouping in group-by clauses?
• How to extend the relational query execution engine with new relational predict operators to answer query evaluation pipelines that include the ML- and LLM-enabled semantics predicates?
• How to optimize queries and efficiently execute these semantic SQL queries and develop new and effective query optimization techniques?
• What are the important research challenges in this domain?

We will give a demonstration of one example Database-for-AI system (iPDB) and how it addresses the above challenges.

Part II — AI for Databases

Here, the target is to leverage ML and LLM techniques in replacing the traditional database components, e.g., query schedulers and query optimizers. We will address the following challenges in the context of two scenarios: (1) ML-based learned data indexes and (2) LLM-based query execution scheduling. More specifically, we will cover the following topics:

• How to replace traditional database components by ML techniques with their "learned counterparts"? And what are the challenges?
• Scenario 1 — ML-based learned data indexes:
  • How ML techniques can replace or augment traditional indexes
  • One-dimensional learned data indexes (both static and dynamic)
  • Multi-dimensional learned data indexes
• How can LLMs be beneficial to databases?
• Scenario 2 — NUMA scheduling using next token prediction:
  • How to schedule main-memory indexes over NUMA servers?
  • How to address the very high numbers of NUMA configurations?
  • How to generalize to unseen hardware and outperform tailored models?
• Research challenges and directions:
  • With ML and LLM models being infused into database engines to optimize for various tasks, e.g., query optimization, performance tuning, data partitioning, scheduling, etc., how to avoid having the database engine being a zoo of ML models of all sorts of types?
  • Generative database agents

Shreya Shankar — Optimizing AI-Powered Data Processing

Carnegie Mellon University

Large language models (LLMs) are increasingly being used not just as chatbots, but as operators inside data processing pipelines at scale, e.g., filtering documents, extracting fields, joining messy records, summarizing evidence. The focus of this workshop is: how do we express large-scale AI computations clearly, and how do we execute them efficiently?

We will begin by reviewing how popular frameworks like MapReduce and SQL decompose computation over collections of data records, then discuss why database systems need query optimization: the same computation can often be executed in many different ways, with dramatically different costs. Students will see how databases represent computations as query plans, how simple choices like pushing down filters can improve performance, and generally why optimization is central to making database systems practical.

We will then introduce a new class of AI operators that apply LLMs to unstructured data, but also create new challenges: LLM calls can be slow, expensive, and imperfect. We will discuss how to optimize query plans that contain AI operators. Students will reason through examples inspired by recent research on AI filters, e.g., using a cheaper model for easy records and a stronger model for harder records, or filtering out irrelevant tokens before sending records to LLMs. The workshop will include short group discussions and paper-based exercises.

Amr El Abbadi — Fault-tolerant Distributed Systems & Private Data Management

University of California, Santa Barbara

Part I — Fault-tolerant Large Scale Distributed Data Management Systems

In this mini-course, we will present some of the fundamental topics that form the basis of research in fault-tolerant distributed systems and computing. The goal is to get up to speed on the foundations of replication as a method of fault-tolerance. This will start by introducing the notion of failures, including crash and malicious faults. Initially the focus is on crash failures and how to achieve consensus in such systems. We will introduce the notion of quorums, and develop the foundation of Paxos, a protocol for managing a replicated state machine using a replicated log. Depending on time, we will then extend consensus protocols developed for crash failures to handle malicious failures.

Part II — Private Data Management: An Introduction

Increasingly, countries and regions have strict laws and regulations to protect the privacy of personal data — for example, the states of the European Union (EU) enforce the General Data Protection Regulations (GDPR) to protect personal data of individuals living in the EU. Much research has focused on preserving the privacy of data using various advanced cryptographic techniques. However, irrespective of the privacy of the data itself, just the queries requesting the data raise severe privacy concerns owing to numerous attacks and data breaches using access patterns.

Our goal in this talk is to survey the basic techniques used to provide private access of data, and then focus on how to use sophisticated, expensive but secure cryptographic methods. We will lay the foundations for large scale privacy preserving, efficient and expressible querying of large data sets using Fully Homomorphic Encryption (FHE). Then we will discuss how to protect the privacy of user queries/tokens in the context of large language models (LLMs). Despite its promise, private FHE presents substantial technical challenges — we will discuss recent efforts to overcome these limitations to improve the performance, scalability and expressiveness of privacy preserving queries of public data and LLM inference.

Jean Hilaire Yapi — Sécurité Oracle DB & Oracle Cloud

Yapson Technology

This hands-on course provides a practical introduction to securing modern Oracle environments. Through a realistic laboratory built on Oracle Database 19c and Node.js, participants learn how to implement auditing, security monitoring, access control, threat detection, and encrypted communications using TCPS.

Designed around real-world scenarios and industry best practices, the course enables participants to develop practical skills in Oracle database security, cloud technologies, and secure application architectures that can be immediately applied in enterprise environments.

Biography

Jean Hilaire Yapi is a Principal Oracle Cloud Architect, Technical Director, and Digital Transformation Expert with over 30 years of experience in enterprise IT. Holding a Ph.D. in Computer Science and an MBA from HEC Montréal, he has contributed to large-scale Oracle, cloud, and modernization initiatives across the financial, industrial, public sector, and technology industries.

His expertise includes Oracle Cloud Infrastructure (OCI), Exadata, Autonomous Database, Oracle E-Business Suite, high availability, automation, hybrid cloud architectures, and database security. Throughout his career, he has worked with organizations such as Desjardins, CGI, CEZinc (Glencore), SNC-Lavalin, Loto-Québec, CAE, and MHIRJ Aviation, helping them modernize, secure, and optimize mission-critical Oracle platforms.

Known for combining deep technical expertise with a practical and business-oriented perspective, Jean designs and delivers hands-on training that bridges the gap between theory and real-world implementation.

Diako Doffou Jérôme — AIR-Net : Réseau Adaptatif Inspiré du Système Immunitaire pour la Détection Incrémentale des Menaces Cybernétiques

Maître-Assistant du CAMES, ESATIC, Côte d'Ivoire

Résumé

L'évolution rapide des infrastructures numériques, notamment les environnements Internet des objets (IoT) et les systèmes distribués, a considérablement accru la complexité et la dynamique des menaces cybernétiques. Les systèmes de détection d'intrusion traditionnels, reposant sur des signatures statiques ou des approches supervisées, peinent à détecter efficacement les attaques inconnues et à s'adapter aux environnements évolutifs.

Nous proposons AIR-Net (Adaptive Immune Response Network), un modèle de détection de menaces bio-inspiré fondé sur les principes du système immunitaire humain, incluant la reconnaissance, la mémoire, l'apprentissage incrémental et l'adaptation. L'architecture introduit des mécanismes de calcul d'affinité, de mise à jour incrémentale des profils comportementaux, ainsi que des stratégies de régulation de la mémoire telles que la fusion et l'apoptose, permettant de maintenir un équilibre entre plasticité et stabilité.

Une évaluation expérimentale menée sur plusieurs jeux de données de référence, incluant NSL-KDD, UNSW-NB15 et CIC-IDS2017, démontre que AIR-Net atteint des performances élevées en termes de précision, de rappel et de F1-score, tout en conservant une empreinte computationnelle réduite. Les résultats mettent en évidence la capacité du modèle à s'adapter aux environnements dynamiques et à détecter efficacement des comportements anormaux.

Ces travaux confirment le potentiel des systèmes immunitaires artificiels pour la conception de solutions de cybersécurité adaptatives, légères et robustes, particulièrement adaptées aux environnements contraints tels que l'IoT.

Mots-clés : cybersécurité adaptative, systèmes immunitaires artificiels, détection d'intrusion, apprentissage incrémental, Internet des objets, détection d'anomalies.

Biographie

Dr Diako Doffou Jérôme est enseignant-chercheur en informatique, Maître-Assistant du CAMES, spécialisé en intelligence artificielle, cybersécurité et cyberdéfense. Il exerce à l'École Supérieure Africaine des Technologies de l'Information et de la Communication (ESATIC), en Côte d'Ivoire.

Ses travaux de recherche portent principalement sur l'apprentissage automatique, l'apprentissage profond, l'intelligence artificielle explicable, la détection des intrusions, l'analyse des vulnérabilités logicielles, la blockchain, les architectures Big Data et la transformation numérique.

Il assure également des responsabilités académiques, notamment dans la coordination des Masters ERIS et e-BIHAR, l'encadrement de mémoires et de projets de fin d'études, ainsi que dans l'organisation des enseignements, des évaluations et des jurys.

Auteur et coauteur de plusieurs publications scientifiques, il participe à des conférences, à des comités de lecture et à des projets de recherche appliquée. Ses travaux récents portent notamment sur l'utilisation de BERT, GPT et des méthodes d'intelligence artificielle explicable pour la détection et l'analyse des cybermenaces, ainsi que sur le développement du modèle AIR-Net pour la détection incrémentale des attaques informatiques.

Mariame Diakité

Fondatrice & Directrice Générale de DIMAN — Leader visionnaire | Experte en Management Organisationnel, Développement d'Affaires, Énergie et Technologies

Biographie

Ingénieure de formation, dirigeante accomplie et entrepreneure engagée, elle cumule plus de trente-huit années d'expérience au plus haut niveau dans les secteurs de l'énergie, des technologies, et du management stratégique. Tout au long de son parcours, elle a contribué à la croissance et à la transformation d'entreprises de premier plan, tout en portant une vision forte : celle d'une Afrique portée par l'excellence et le développement durable.

Diplômée de l'École Nationale des Ingénieurs de Spécialités (ENIS-INSET de Yamoussoukro, aujourd'hui intégrée à l'INP-HB), Mariame Diakité est titulaire d'un diplôme d'Ingénieur en Électronique.

Animée par une volonté constante d'apprentissage et de perfectionnement, elle a enrichi son parcours par plusieurs formations exécutives de renommée internationale, notamment l'Advanced Management Program (AMP) de MDE Business School, dont elle a assuré la présidence de promotion.

Elle a également suivi le programme ASPIRE, développé par African Management Institute (AMI) en collaboration avec la Stanford Graduate School of Business.

Elle a en outre renforcé ses compétences avec l'obtention d'un Certificat d'Aptitude au Métier de Consultant (CAMEC) auprès du CCDE.

Depuis 2021, elle dirige DIMAN, un Cabinet spécialisé dans le conseil, les études et les solutions technologiques. À travers cette structure, elle accompagne des groupes de référence dans leurs projets de transformation et de développement.

Elle représente localement, dans le cadre de son déploiement de solutions de comptage intelligent à grande échelle, le Groupe WASION, un des principaux fournisseurs mondiaux de solutions et d'équipements intelligents pour la gestion et le comptage de l'énergie (électricité, eau, gaz).

Auparavant, en qualité de Directrice Générale du Groupe KAMA, sous son leadership, d'importants projets ont été développés, notamment dans les domaines du gaz et du transport d'hydrocarbures.

Entre 2003 et 2016, au sein de Schneider Electric, leader mondial des technologies de l'énergie et des automatismes, elle a également occupé la fonction de Directrice Générale de Schneider Electric Afrique de l'Ouest Francophone. Elle a contribué à la réalisation de nombreux projets avec les Régies d'Électricité et les Industries, et au développement en Afrique de l'Ouest d'un dynamique réseau de partenaires locaux (distributeurs, installateurs, systèmes intégrateurs).

Au-delà de ses responsabilités de dirigeante, Mariame Diakité est profondément attachée à la transmission du savoir et à l'accompagnement des nouvelles générations.

Elle entretient depuis toujours un lien privilégié avec son alma mater, l'INP-HB, où elle intervient auprès des étudiants et soutient diverses initiatives académiques. Engagée en faveur du développement du leadership féminin et de la promotion des compétences africaines, elle exerce aujourd'hui plusieurs fonctions de conseil et de mentorat, notamment au sein de l'Ordre des Ingénieurs Industriels de Côte d'Ivoire (OIICI) et de l'Association des Professionnels des Énergies Renouvelables de Côte d'Ivoire (APERCI).

Par son parcours, son sens du leadership et sa capacité à transformer les défis en opportunités, Mariame Diakité incarne l'excellence et l'audace.