The course aims at providing PhD Candidates knowledge on basic state of the art and advanced theories/techniques for learning from multisensory signals and data Bayesian models for jointly predicting, processing, filtering and interpreting observed interactions.

Such models will be shown to enhance the functionalities of embodied smart autonomous systems like cars, radios, drones, robots, buildings, by providing such agents with a self-awareness information layer. Networks of self-aware autonomous systems interacting in smart cognitive environment will be also targeted as examples carried on in the course. From a methodological viewpoint, this module aims at identifying and describing methodologies and techniques for defining a common probabilistic framework suitable for:

• integrating contextual signals synchronously provided by multisensorial exo and proprio receptive sensors of autonomous systems by using Data Fusion paradigms and techniques;
• learning from experiences behavioural and causal self-awareness models allowing an autonomous system to describe the world through a vocabulary of normal locally stationary experiences.
• showing how each model learned from experience can describe through probabilistic stationary rules dynamic perception, planning and actuation by means of collected external and internal observations.

Applications will be targeted of described techniques related to a couple of main case studies together with additional examples:

• self-awareness in autonomous ground and aerial vehicles and smart infrastructures (e.g., Buildings, dynamic radio spectrum)
• interactions in telecommunications scenarios like cognitive radio and internet of things.

Program:

• Bio-inspired Cognitive Dynamic System models: Damasio, Haykin and Friston models
• Data Fusion methodologies and techniques for integrating multisensorial contextual data Data Fusion models. JDL model and its extensions: signals, objects, situations, threats, processes and cognitive refinement.
• Techniques for non-parametric self-awareness interaction-based predictive/generative/classification models.
• Dynamic Bayesian Networks as representation and inference tool for self-awareness
• Review of Bayesian multisensor state estimation and data association techniques:
  • Continuous and discrete state estimation techniques: from Kalman filter to Particle Filters, Hidden Markov Models.
   
  • Switching models. Markov Jump and Rao Blackwellized filters.
  • Relationships with Generative Adversarial Networks, Variational Autoencoders.
• Supervised and Unsupervised Learning of DBNs models from multisensorial agent exo and proprioreceptive sequences:
  • Gaussian Processes,
  • Self Organizing Maps, Growing Neural Gas.
• Incremental learning: Dirichlet processes, DBN error based new model creation
• Applications and case studies:
  • Cognitive radio and Internet of Things Physical anti-jammer security Self-awareness in autonomous ground and aerial vehicles
  • Cognitive safety and physical security systems (smart patrolling in cooperative environments, preventive automotive vehicles, smart buildings, etc.)
• The course includes two laboratory lessons. The objective of these lessons is to get familiar with the application of generative models to low-dimensional data. We will explore how Dynamic Bayesian Networks (DBNs) can be utilized for anomaly detection considering single objects, and how interactions among multiple objects can be modeled using coupled DBNs. This allows participants to connect concepts learned during theoretical lessons with practical implementation capabilities.

Enrollment form (it is mandatory to enroll through this form before 03/07/2023)