Lectures

Lecture 1/3 “Introduction to Multi-Agent Reinforcement Learning”

Abstract TBA

Lecture 2/3 “If Multi-Agent Foundation Models is the Answer, What is the Question?”

Abstract TBA

Lecture 3/3 “Becoming an AI Researcher: Practical Advice for Graduate Students”

Abstract TBA

Lecture 1/3 “Predictive coding: training neural networks with local rules”

Abstract TBA

Lecture 2/3 “Predictive coding: advantages over backpropagation”

Abstract TBA

Lecture 3/3 “Predictive coding: implementation”

Abstract TBA

Lecture 1/3 “Foundations of Guardrailing Agentic AI via Temporal Synthesis”

ABSTRACT: This course presents temporal synthesis as a foundation for guardrailing agentic AI. Temporal synthesis studies the automatic synthesis of interactive programs (strategies) from declarative specifications expressed in temporal logic. In this course, we show how temporal synthesis provides a practical foundation for strategic reasoning in autonomous AI systems. The key to this research path lies in the rich body of concepts developed in reasoning about actions and planning, combined with a precise treatment of nondeterministic environments and temporal objectives. In such settings, plans must be treated as strategies rather than being blurred with individual execution traces, and goal satisfaction evolves during execution rather than being reducible to reaching states with fixed properties, as in classical planning. These features are naturally captured within the temporal synthesis framework. Technically, we study the structure of the entire space of strategies satisfying a specification. Rather than assuming strategies to be observable, we focus on characterizing the set of execution traces that are compliant with a given strategy space. This perspective provides formal support for guardrailing and responsibility attribution in systems composed of multiple, independently acting autonomous agents.

Lecture 2/3 “Foundations of Guardrailing Agentic AI via Temporal Synthesis”

ABSTRACT: This course presents temporal synthesis as a foundation for guardrailing agentic AI. Temporal synthesis studies the automatic synthesis of interactive programs (strategies) from declarative specifications expressed in temporal logic. In this course, we show how temporal synthesis provides a practical foundation for strategic reasoning in autonomous AI systems. The key to this research path lies in the rich body of concepts developed in reasoning about actions and planning, combined with a precise treatment of nondeterministic environments and temporal objectives. In such settings, plans must be treated as strategies rather than being blurred with individual execution traces, and goal satisfaction evolves during execution rather than being reducible to reaching states with fixed properties, as in classical planning. These features are naturally captured within the temporal synthesis framework. Technically, we study the structure of the entire space of strategies satisfying a specification. Rather than assuming strategies to be observable, we focus on characterizing the set of execution traces that are compliant with a given strategy space. This perspective provides formal support for guardrailing and responsibility attribution in systems composed of multiple, independently acting autonomous agents.

Lecture 3/3 “Foundations of Guardrailing Agentic AI via Temporal Synthesis”

ABSTRACT: This course presents temporal synthesis as a foundation for guardrailing agentic AI. Temporal synthesis studies the automatic synthesis of interactive programs (strategies) from declarative specifications expressed in temporal logic. In this course, we show how temporal synthesis provides a practical foundation for strategic reasoning in autonomous AI systems. The key to this research path lies in the rich body of concepts developed in reasoning about actions and planning, combined with a precise treatment of nondeterministic environments and temporal objectives. In such settings, plans must be treated as strategies rather than being blurred with individual execution traces, and goal satisfaction evolves during execution rather than being reducible to reaching states with fixed properties, as in classical planning. These features are naturally captured within the temporal synthesis framework. Technically, we study the structure of the entire space of strategies satisfying a specification. Rather than assuming strategies to be observable, we focus on characterizing the set of execution traces that are compliant with a given strategy space. This perspective provides formal support for guardrailing and responsibility attribution in systems composed of multiple, independently acting autonomous agents.

Lecture 1/3 “Computational Cognitive Modeling Using Reinforcement Learning”

Abstract TBA

Lecture 2/3 “Advancing Cognitive Modeling by Combining Hand-Crafted and Neural-Network Models”

Abstract TBA

Lecture 3/3 “Advancing Cognitive Modeling using Procedurally-Generated Task Spaces”

Abstract TBA

Lecture 1/3 “On the impact of data for vision-language learning (1/2)”

Abstract TBA

Lecture 2/3 “On the impact of data for vision-language learning (2/2)”

Abstract TBA

Lecture 3/3 “Building frontier visual segmentation models”

Abstract TBA

Lecture “Agentic AI part 1”

Abstract TBA

Lecture “Agentic AI part 2”

Abstract TBA

Lecture “Agentic AI part 3”

Abstract TBA

Lecture 1/2 “Agentic Security (how to attack and defend against prompt injection)”

Abstract TBA

Lecture 2/2 “Agentic Security (how to attack and defend against prompt injection)”

Abstract TBA

Lecture “Generative Smart Spaces – A Novel Physical-AI Approach to Programming Future Internet-of-Things”

The Internet of Things has inspired compelling visions of intelligent homes, hospitals, cities, and environments. Yet a fundamental question remains largely unresolved: How do we actually program the IoT? Unlike conventional computers or cloud applications, the IoT is a dynamic computational environment whose components sense and act on the physical world, appear and disappear, expose different capabilities, and form opportune interactions that may not have been anticipated by their developers.

Our early work approached this problem from the bottom up. Service-Oriented Device Architecture (SODA) and the Atlas platform and middleware sought to make the Thing an explicit, self-describing, and programmable computational entity. This enabled service composition in smart spaces and motivated later work on self- and peer-conscious Things, collaborative microservice programmability, and transactional extensions for safer cyber-physical execution (IoTranx). The central lesson is that there are no shortcuts to IoT programmability: the architecture, semantics, and capabilities of Things themselves matter.

But making Things programmable is only half of the challenge. Future IoT environments will be too large, dynamic, and opportunistic for developers to manually anticipate every possible interaction. The next step is therefore to create IoT systems that can generate, organize, and understand their own interaction spaces.

After a quick preview of our SODA and Atlas journey, I will present two emerging directions toward this vision. The first is Generative MQTT (gMQTT), which introduces a generative broker-control plane capable of synthesizing and evolving MQTT-compatible topic spaces from semantic device descriptions, runtime context, user interactions, and/or application goals. Rather than assuming that topics, subscriptions, and routing structures are known in advance, gMQTT generates inspectable and verifiable broker configurations. Large language models serve a constrained semantic mediation role rather than directly controlling the physical environment.

The second direction, Sensor Language Modeling, asks whether the sensors of an intelligent environment may collectively be speaking their own language! Instead of treating sensor data only as numerical signals mapped to predefined activities, we view sensor events as words, temporal combinations as phrases and sentences, and longer sequences as narratives of the evolving smart space. By combining static device semantics with dynamic meanings learned from sensor behavior, foundation models may discover the language of intelligent environments.

As a Conclusion, we show that, together, these directions point toward a future in which we explicitly architect the Things, while enabling the IoT to generate its interaction structures, learn the language of its environment, and ultimately be programmed at the level of intent.

Lecture 1/3 “The rise of Tabular Foundation Models”

Abstract TBA

Lecture 2/3 “TabICL: on open Tabular Foundation Model”

Abstract TBA

Lecture 3/3 “Next frontiers for TFMs”

Abstract TBA

Lecture 1/3 “The rise of generative AI agents: reasoning, memory and planning capabilities”

Abstract TBA

Lecture 2/3 “The social intelligence of generative AI agents: social norms, social reasoning and Theory of Mind”

Abstract TBA

Lecture 3/3 “A society of generative AI agents: emergence of collective behaviors, safety and implications for human societies”

Abstract TBA

Lecture 1/3 “Neuroevolution: The basics”

Abstract TBA

Lecture 2/3 “Neuroevolution: Diversity, Novelty, and Open-Ended Discovery”

Abstract TBA

Lecture 3/3 “Neuroevolution: Synergies with Generative AI”

Abstract TBA

Lecture 1/3 “Complex Event Recognition/Forecasting and Applications”

Complex Event Recognition (CER) aims to identify complex, time-evolving situations of interest by detecting temporal and relational patterns in continuously arriving event streams, while Complex Event Forecasting (CEF) seeks to anticipate whether and when such situations will unfold, from early signs. In this talk we introduce the fundamental concepts and challenges of CER/F and present formal approaches based on computational logic and temporal formalisms, such as the Event Calculus and symbolic automata. We will examine how complex event patterns are represented, how they are efficiently matched against high-volume streams, and how they can be combined with probabilistic sequence models to produce forecasts and estimates of their expected occurrence time. The methods will be illustrated through applications such as maritime situational awareness, human activity recognition, transport/mobility monitoring and personalized medicine, demonstrating how CER/F transforms heterogeneous streaming data into timely, actionable knowledge.

Lecture 2/3 “Large-Scale Probabilistic Complex Event Recognition/Forecasting”

Real-world event streams are characterized by noisy perception and epistimic uncertainty, making purely crisp event inference insufficient. In this talk we will present probabilistic approaches to CER/F that associate confidence values with both incoming events and the complex situations inferred from them. We will discuss probabilistic extensions of logical temporal models and explain how possible-world inference can be implemented efficiently through knowledge compilation and weighted model counting. By compiling recurring reasoning structures into reusable computational circuits and by expressing temporal inference through parallel tensor operations, these approaches retain declarative temporal semantics while supporting scalable online data processing under uncertainty. We will also present probabilistic forecasting methods that combine symbolic event patterns with learned stochastic sequence models to estimate both the likelihood and the expected timing of future complex events.

Lecture 3/3 “Neuro-Symbolic Reasoning and Learning for Complex Event Recognition/Forecasting”

In the last part of the course we will explore how symbolic temporal reasoning and neural learning can be combined to develop CER/F systems that learn from raw data while retaining structured, interpretable models of complex situations. First, we will consider architectures in which neural perception modules detect low-level events from video or sensor streams, while differentiable temporal logic models compose them into complex events. In this setting, the symbolic model acts as a “teacher” for neural perception: supervision available only for high-level complex events can propagate through temporal reasoning to train the underlying perception modules. We will then move beyond predefined event patterns and present neuro-symbolic methods for learning their structure from labelled sequences, including the induction of executable temporal rules and symbolic automata. These approaches combine the adaptability of neural learning with the compositionality, explainability and formal semantics of symbolic CER/F. Finally, we will discuss neuro-symbolic formal verification: establishing whether perturbations of the raw input can propagate through both the neural and symbolic components and alter system-level conclusions or violate safety properties. We will present scalable relaxation and bound-propagation techniques that can provide sound robustness certificates for the complete perception-reasoning pipeline.

Impact of AI and Optimization on the economics of sustainability