Goal
Upon successful completion of the course, the student will be able to:
- Understand the complexity of the basic methods of artificial intelligence (inference, machine learning), combined with the needs of modern applications, which require the management of volumes of data at multiple scales
- Choose the most appropriate AI method (and also data pre-processing) according to the requirements of the given problem, based on the volume of available data and the abundance (or scarcity) of computing resources
- Understand and identify the sources of error and biased decisions of an AI system, as well as provide a quantified evaluation of it in terms of the volume of data it requires and the effects that increasing/decreasing the volume has on accuracy, (computational) cost, reliability and its interpretability
- Understand and use artificial intelligence formalisms that enable human-machine synergy to solve artificial intelligence problems, through the construction of interpretable, verifiable and reliable machine learning models (interpretable vs. black-box models)
Also, the course targets to the following general competencies:
- Ability to organize and plan work and manage time effectively
- Ability to communicate effectively (orally and written)
- Ability to solve problems
- Ability to develop critical thinking and capacity for critical approaches
- Ability to work in a team
- Ability to apply theoretical knowledge in practice
Contents
The recent developments in the field of generative AI (Generative AI - GenAI) have highlighted, more strongly than ever, concerns regarding the role of scalability in the development of AI systems, but also the reliability of such systems. On the one hand the huge sizes of the training data and the multitude of parameters are largely responsible for the recent impressive successes in AI.
On the other hand, these approaches show serious reliability problems, which significantly limit their usefulness in critical applications. The aim of the course is to familiarize the students with the relevant concerns and the current debate about the role of scalability and reliability in ΑΙ. At the same time, the familiarity with "hybrid" ΑΙ techniques, which seek to combine methods from different areas of ΑΙ, with the aim of developing reliable ΑΙ systems, scalable at multiple levels (e.g. both large and small data sets).
These techniques run through almost the entire AI "arsenal", including (deep) machine learning, knowledge representation and inference, optimization and formal methods. Therefore, the course also offers an overview of these different fields of AI and their combinations in applications with particular scalability and reliability requirements.
Course content:
- Introduction: scalability and reliability problems in AI. How far can we go with "big data" and how can we trust the behavior of an AI system trained on it? Trustworthy AI and its international standardization. Transparency, robustness, reliability, fairness etc. Overview of complex ΑΙ systems and applications with increased scalability and reliability requirements (eg autonomous systems, medical decision support systems, etc.)
- Scalability in large volumes of data (scale-out, scale-up). Real-time AI methods. Big Data manipulation and pattern recognition in data streams. Distributed pattern recognition. Handling concept drift
- Scalability in small volumes of data (scale-in) and in conditions of scarce computing resources. Introduction to frugal machine learning (Frugal ML). Evaluation of AI methods considering limitations in the amount of available data and computing resources
- Overview of symbolic techniques. Logic, knowledge representation and inference, formal languages and automata. SAT-based techniques and combinatorial optimization, standard methods. Scalability of symbolic techniques and their role in Trusted AI.]
- Interpretability of models (interpretable ML) and explainability of predictions (explainable AI - XAI). Overview of basic XAI methods (LIME, SHAP, logic-based), examples and applications. Scalability of explanatory techniques
- Overview of machine learning (neural, statistical, symbolic) and scalability and reliability issues. Robustness issues in (deep) machine learning. Formal verification of neural models. Robustification through training with verification counterexamples
- Techniques for combining learning and logical-probabilistic inference, Introduction to neuro-symbolic AI. Algebraic view of logical inference and differentiable inference. End-to-end training of neural models with symbolic knowledge
- Neuro-symbolic techniques for learning from less data and imposing plausible behavioral constraints on neural models. Examples of functional constraints on correctness, security, and impartiality
- Neuro-symbolic techniques and Generative AI. Scalability, interpretability and correct behavior
- Standard verification of neuro-symbolic models and complex systems. From individual component reliability to system reliability. Standard verification of high-level properties (verification of system-level properties)
Bibliography
- Kaur, D., Uslu, S., Rittichier, K. J., & Durresi, A., Trustworthy Artificial Intelligence: A Review. ACM computing surveys (CSUR), 2022
- Darwiche, A. Three Modern Roles for Logic in AI. ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, 2020.
- Marques-Silva J. and Ignatiev A., Delivering Trustworthy AI through Formal XAI, AAAI, 2022
- Albarghouthi, A. Introduction to neural network verification. Foundations and Trends in Programming Languages, 2021.
- Hitzler, P., & Sarker, M. K. (Eds.), Neuro-Symbolic Artificial Intelligence: The State of the Art, IOS Press 2022
- Marra, G., Dumančić, S., Manhaeve, R., & De Raedt, L. From statistical relational to neurosymbolic artificial intelligence: A survey. Artificial Intelligence, 2024.
- Relevant scientific journals:
- Journal of Artificial Intelligence Research (JAIR), ISSN: 1076-9757
- Journal of Machine Learning Research (JMLR), ISSN 1533-7928
- Artificial Intelligence, Springer, ISSN: 0269-2821
- Neurosymbolic Artificial Intelligence, 2949-8732
- Machine Learning, Springer, ISSN: 0885-6125
- Relevant congresses:
- All conferences on artificial intelligence and machine learning: AAAI, IJCAI, KR, ECAI, ICML, ECML, NeurIPS.
