Gillain Emmanuel (ed.). Demystifying Artificial Intelligence: Symbolic, Data-Driven, Statistical and Ethical AI

De Gruyter, 2024. — 476 p. — ISBN: 978-3-11-142567-2.This book is intended for business professionals who want to understand the fundamental concepts of Artificial Intelligence (AI), their applications, and limitations. Built as a collaborative effort between academia and the industry, this book bridges the gap between theory and business application, demystifying AI through fundamental concepts and industry examples. The reader will find here an overview of the different AI techniques to search, plan, reason, learn, adapt, understand, and interact. The book covers the two traditional paradigms in AI: the statistical and data-driven AI systems, which learn and perform by ingesting millions of data points into Machine Learning algorithms, and the consciously modeled AI systems, known as symbolic AI systems, which use explicit symbols to represent the world and make conclusions. Rather than opposing those two paradigms, the book will also show how those different fields can complement each other.Out of the different fields of AI, much of the attention over the last years has been focused on the field of Machine Learning (ML), its subfield of artificial neural networks with Deep Learning, and natural language processing (NLP).Developing performant traditional Machine Learning systems requires a pipeline of tasks with multiple choices and fine-tuning decisions to achieve the most optimal result (the predictors, the model itself with configuration and hyperparameters, etc.). The search space to find the optimal parameters is sometimes complex with multiple dimensions so mathematicians would classify the data scientist’s tasks as “high-dimensional combinatorial optimization” tasks. Automated machine learning (AutoML) is an idea that emerged in the 1990s and whose objective is essentially to automate the generation and selection of the most performing algorithms and optimize their performance without the help of data scientists: a data scientist applying AutoML techniques would typically only require a couple of lines of codes to test multiple models with multiple hyperparameters in parallel, and let the algorithm select the best model under some defined quality metrics. These methods have progressively been making their way into standard commercial products as a productivity tool that helps data scientists work faster and better. AutoML techniques speed up the model development cycle, often with more performant models.Supported by an ever-increasing amount of data and processing power to train the algorithms, those research efforts resulted in a series of breakthroughs that can be illustrated by impressive progress made across very different fields like image classification, object detection, or image generation in computer vision, machine translation, the creation or recognition of speech, text understanding and writing, complex board games, etc.Preface and introduction.
A holistic view of AI techniques, their limitations, and complementarities.
Solve problems by searching, including with constraints, a fundamental pillar.
Reasoning with first-order logic.
Knowledge representation and engineering with ontologies.
Probabilistic reasoning: When the environment is uncertain.
Learning from data.
Between language and knowledge.
Some words about ethics. The angles of fairness and transparency.
Industry examples where different AI techniques are combined.
Conclusion – Moving forward.