Tutorials

Abstract

Every digital twin of an intelligent environment must have an internal knowledge representation model, which collects the information about the physical environment, its current state, its potential behavior, historical states and potential what-if situations and their constraints, etc. The representation layer includes simulation capabilities, modelling paradigms, reasoning, interoperability with other applications, and other “intelligent” capabilities of the system. Yet the form of representations is often decided during the design stage of the digital twin system, and largely made implicitly, based on the experience of the architects and available tools, rather than by considering all available options. This tutorial gives a foundational structured overview of the four families of representation models found in digital twins and cyber-physical systems: mechanistic models grounded in natural laws, data-driven models learned from data, logic-based formal models that specify and verify behavior, and ontology-based models that provide interoperable, human-understandable vocabularies. The tutorial describes each family in detail, discussing its assumptions, requirements, costs to make and to run, and discusses, how other simulation techniques depend on the chosen representation model. The tutorial also shows that the most valuable results often arise from combining the representation paradigms, and discusses diQerent ways of achieving it, such as surrogate models, physics-informed neural networks, rule learning and symbolic regression, formalization and schema validation, and foundation models. Illustrative examples follow each family of representations and their combinations.

Abstract

Modern intelligent following of environmental influences increasingly rely on heterogeneous sensors, edge computing, artificial intelligence, and context-aware decision-making to understand people and their surroundings and to respond to potentially dangerous situations. This tutorial introduces the principles and practical challenges of designing multimodal, edge-AI intelligent environments for health and safety in high-risk conditions, using the NATO SPS PRIME Alert – Personalised Micro Environmental Change Alert project as a real-world case study.

The tutorial will guide participants through the complete design chain, from physiological and environmental sensing to multimodal data fusion, local AI inference, risk estimation, personalised alerts, and command-level decision support. Particular attention will be given to the integration of vital signs, speech and micro-environmental measurements information; the division of intelligence between wearable/mobile devices, edge modules, and cloud infrastructure; and the challenges of operating under unreliable connectivity, limited computational resources, and rapidly changing environmental conditions. Through examples participants will explore how AI can transform heterogeneous sensor streams into actionable information while addressing reliability, explainability, privacy, human factors, and real-world deployment constraints.

Abstract

Intelligent Environments (IE), such as smart homes, ambient assisted living facilities, and IoT-enabled smart cities, generate large volumes of heterogeneous sensor and interaction data. Understanding the underlying patterns within these spaces requires machine learning models that are not just accurate, but highly interpretable for domain experts. Subgroup Discovery (SD) is a descriptive and supervised machine learning technique specifically designed to find subsets of data with statistically interesting characteristics regarding a target variable of interest (e.g., identifying a specific combination of sensor activations that precede an anomalous behaviour in a smart home).

This tutorial introduces ‘Subgroups’, a comprehensive and open-source Python library developed to streamline SD tasks. Through this hands-on tutorial, attendees will learn the theoretical foundations of SD and practically apply the library to IE datasets. By the end of the tutorial, participants will be able to generate human-readable patterns that explain intelligent environments.

Abstract

Foundation models are increasingly used for the interpretation of sensor data, including physiological signals such as PPG and ECG, as well as inertial signals commonly acquired from wearable devices. They are typically trained on unlabelled data to learn rich latent representations capturing interesting physiological or behavioural phenomena. They output embeddings that can be used in further processing, or the models can be fine-tuned for particular tasks. Since the foundation models are usually trained on considerably more data than researchers and practitioners typically have available for their tasks, their representations can be more robust and general than those obtained by training on task-specific data. This can translate to better task performance on varied downstream tasks, especially after fine-tuning. The clearest example of this are large language models (LLMs), which are also a type of foundation model. Some prominent examples for human sensing are Google’s LSM, which was trained on 40 million hours of multimodal wearable data from 165,000 participants, and PaPaGei, which was trained on 57,000 hours of PPG data. The tutorial will present the general principles of foundation models for human sensing, as well as a broad range of existing models.

Another topic that will be addressed by the tutorial is the combination of sensor models with LLMs. This can be done in ways analogous to multimodal vision-language models, for example by representing sensor data and language in a joint latent space or by providing language-based descriptions of sensor-derived features. This enables conversations about sensor data in natural language, which is useful for data exploration and even tasks such as human activity recognition.  Finally, the tutorial will tackle the practical question of how to adapt a foundation model to one’s task. We will use PPG and blood pressure estimation as a case study, presenting the relevant foundation models, datasets and fine-tuning methods. We will also explain how this approach can be generalised to other models and tasks. For the case study, we will use a service developed within SLAIF, designed to provide an accessible way of fine-tuning such models and coupled with a robust evaluation framework.

Abstract

In today’s intelligent work environments and interconnected world, digital forensic investigators face new challenges in ensuring a proper digital forensic process. Although the number of cybersecurity strategies, policies, and protocols has increased significantly in recent years, the cyberattacks have simultaneously continued to grow. Despite the gain of research and tools in the domain, the meaning of digital forensics for cybersecurity has seen limited discussion. Digital forensics is defined as “the process of identifying, preserving, analyzing, and presenting digital evidence“. This tutorial offers a practical, hands-on introduction to digital forensics applied to a simulated incident in a smart office. Participants will use industry-standard tools – FTK Imager, Autopsy, and HexEdit to reconstruct an event timeline, recover deleted and deliberately disguised files, and correlate digital traces from multiple sources, including access logs, sensor data, and workstation artifacts through a realistic scenario involving unauthorized after-hours access to a server room and an attempted exfiltration of financial data. The session concludes with a discussion of the specific technical challenges of IoT/edge forensics and the accompanying ethical and legal implications for evidence collection in intelligent workplaces.

Abstract

Foundation-model benchmarks increasingly shape which models get selected, adapted, and deployed — including into the sensor-rich, resource-constrained, and often multilingual settings that define intelligent environments. Yet benchmark outcomes are highly sensitive to evaluation design choices such as dataset composition, aggregation methodology, and metric weighting, which can produce misleading conclusions about model capability and superiority. This tutorial presents recent advances in robust, transparent, and evidence-based evaluation of foundation models: ranking robustness, sensitivity to dataset composition, benchmark redundancy, multilingual and low-resource evaluation, and multi-criteria model assessment under conflicting objectives (accuracy, efficiency, sustainability, fairness). Through case studies on text embedding models, large language models, and tabular foundation models, attendees will see how benchmark conclusions can change under different settings, and will gain practical tools for selecting models and designing trustworthy evaluation pipelines for deployment in intelligent environments.

Abstract

AI-driven laboratories, sometimes called self-driving or autonomous labs, close the loop between experiment and analysis: an algorithm selects the next experiment, a robotic system executes it, the results are analyzed, and the cycle repeats without human intervention. These systems are intelligent environments in the most literal sense, and they are rapidly transforming chemistry, materials science, and biology. This tutorial teaches the two methods at the core of most AI-driven labs: Gaussian process (GP) regression, which provides predictions with calibrated uncertainty from small, expensive datasets, and Bayesian experimental design, which uses those uncertainties to decide what to measure next. Roughly half of the session is hands-on. Participants will work through Google Colab notebooks that demonstrate GP use on real world data, explore acquisition functions, and run a complete simulated autonomous experimental campaign. No prior experience with GPs or laboratory automation is required. Participants leave with working code, an understanding of the design decisions that matter in practice, and a clear picture of how to apply these methods to their own sensing, control, and optimization problems.