OBIETTIVI FORMATIVI

The goal of this module is to provide students with knowledge and capabilities for processing heterogeneous multisensory signals acquired by autonomous and semiautonomous systems, including human machine interaction. Machine learning methods will be addressed for estimating optimized Bayesian dynamic cognitive models in a data driven way. Capabilities of the students to produce new solutions based on such methods to state of the art problems will be addressed, together with acquisition of programming related capabilities by means of laboratory activities.

OBIETTIVI FORMATIVI (DETTAGLIO) E RISULTATI DI APPRENDIMENTO

 La frequenza e la partecipazione attiva alle attività formative proposte (lezioni frontali e attività in laboratorio) e lo studio individuale consentiranno allo studente di

-Conoscere i fondamenti sulle architetture e le logiche di rappresentazione della conoscenza in  sistemi e reti di agenti cognitivi autonomi e semi autonomi multisensoriali, artificiali, dinamici ed autoorganizzanti

- Conoscere  le principali metodologie probabilistiche di rappresentazione , elaborazione e fusione di segnali temporali propriorecettivi ed exterorecettivi multisensoriali in agenti cognitivi dinamici (p.e. veicoli semi autonomi, radio cognitive, etc.) 

- Conoscere  le metodologie di apprendimento automatico data driven e non supervisionato di modelli dinamici predittivi e di analisi (modelli generativi) a partire da sequenze temporali  di segnali eterogenei

-Conoscere e acquisire  capacità di uso di strumenti di programmazione e di simulazione per analizzare dataset preregistrati di sequenze temporali ed applicare a tali sequenze metodi di apprendimento di modelli generativi e metodi di predizione e stima dinamica dello stato degli agenti.

- Integrare le conoscenze suddette nell'ambito di framework di laboratorio per sviluppare prototipi simulativi di agenti cognitivi dinamici artificiali con funzionalità di la Situation awareness e di Self awareness 

- Acquisire la capacità di analizzare, commentare  e descrivere i risultati  ottenuti applicando quanto appreso a dataset simulati o ricavati da sistemi  artificiali che descrivano situazioni dinamiche di agenti autonomi (p.e.veicoli  autonomi, robot e droni), sotto forma di report, presentazioni con slide,poster,

PREREQUISITI

Essendo un corso magistrale del secondo anno, lo studente deve avere acquisito:

- conoscenze di base sulla teoria delle probabilitità, i processi casuali, la teoria dei segnali.aleatori

- conoscenze di base di programmazione e ambienti di programmazione p.e. Matlab

MODALITA' DIDATTICHE

The course is divided in two parts. Lectures in frontal teaching modality presented together with slides will aim at describing the theroretical concepts and the techniques. Such lectires will cover 40 hours and can be recorded and made available on those channels recommended by Univeristy of Genova. The second part is done within a laboratory carried on by an expert of the field and will involve application of programs in Matlab framework that correspond to theories and techniqeus shon at lectures. Students will be required to present a report at the end of each lab experience. 10 Lab experiences are planned and will help students to be prepared to present the final report to be discussed during the exam that will show application and discussion of techniques abnd results over a dataset assigned.. 

PROGRAMMA/CONTENUTO

​

1. Introduzione ai sistemi cognitivi dinamici
   1. Sensori propriorecettivi e exterorecettivi
   2. Rappresentazione ed inferenza per la self awareness
   3. First person and Third Person representations
   4. Self awareness as result of incremental learning of probabilistic representations
2. Bayesian Networks
   1. General concepts: conditional independence, examples, Markovianity. Inference: Naive method. Belief propagation
   2. Dynamic Bayesian Networks. Slices, model unfolding, Kaman filters and HMM as DBNs . Filters as two level Bayesian networks
3. Bayesian inference: filtering, smoothing, prediction and update
   1. Kalman Filter: prediction and update. Discussion of model filter and example
   2. Hidden Markov Models, Extended Kalman Filter, Unscented Kalman Filter and Cubature Kalman Filter
   3. Other Filters: Information Filter, Square root Filter, H infinity filter
   4. Particle Filter: Non parametric probability representation; SIR and SIS
4. Hierarchical representations and Switching models
   1. Hierarchical DBNs: including discrete and continuous variables in DBN slice hierarchy
   2. Switching models: hierarchical feature representation Data Fusion inference and data equivalence with HDBNs
   3. Markov Jump Particle Filter Rao Blackwellized Particle filter
5. Generative models: representation and inference
   1. HDBN with continuous switching variables.
   2.  Generalized coordinates and generalized filtering
   3. Generalized filtering: free energy minimization as learning paradigm
6. Machine learning for HDBNs
   1. Learning generative models from multisensory sequences
   2. Unsupervised Clustering and HDBN learning
   3. Self organizing Map,
   4. Growing Neural Gas and ITM
   5. Learning conditional probabilities:
      1. parameter learning,
      2. HMM and KF learning methods (EM, Baum Welch, etc.)
      3. Gaussian processes.
      4. Dirichlet process
      5. ELBO variational approach
7. Incremental learning of switching generative models.
   1. Generalized DBNs as result of agent’s self poietic dynamic stability control.
   2. Abnormality detection methods for loss of stability detection: probabilistic distances
      1. Bhattacharrya,
      2. Hellinger
      3.  Mahalanobis
      4. Kullback Liebler.
   3. Null force filter, abnormality detection and generalized errors
   4. Learning new generative models using generalized errors.
8. Application and integration of course concepts by multisensory dataset processing. Assisted Laboratory experiences and report preparation
   1. Introduction to Matlab
   2. Bayesian Networks
   3. Kalman Filter
   4. Particle Filter
   5. Hidden Markov Models
   6. Markov Jump Particle Filter
   7. Generalized Filtering
   8. Unsupervised Clustering
   9. Markov Jump Particle Filter and anomaly detection
   10. Integration of methods for incremental learning
   11. Report presentation and  discussion

TESTI/BIBLIOGRAFIA

 Tutte le slides utilizzate durante le lezioni ed i laboratori e altro materiale didattico saranno disponibili su aul@web. In generale, gli appunti presi durante le lezioni e il materiale su aul@web sono sufficienti per la preparazione dell'esame.Tutte le slides utilizzate durante le lezioni e altro materiale didattico saranno disponibili su aul@web. In generale, gli appunti presi durante le lezioni e il materiale su aul@web sono sufficienti per la preparazione dell'esame.  

- A. R. Damasio, Looking for Spinoza: Joy, Sorrow, and the Feeling Brain, 1st ed. Orlando: Harcourt, 2003. [Online]. Available:http://lccn.loc.gov/2002011347
- S. Haykin, Cognitive Dynamic Systems: Perception-action Cycle, Radar and Radio, ser. Cognitive Dynamic Systems: Perception–action Cycle, Radar, and Radio. Cambridge University Press, 2012.

- P. R. Lewis, M. Platzner, B. Rinner, J. Torresen, and X. Yao, Eds., Self aware Computing Systems: An Engineering Approach. Springer, 2016.

- K. J. Friston, B. Sengupta, and G. Auletta, “Cognitive dynamics: From attractors to active inference,” Proceedings of the IEEE, vol. 102, no. 4, pp. 427–445, 2014. [Online]. Available:
https://doi.org/10.1109/JPROC.2014.2306251

- S. Haykin and J. M. Fuster, “On cognitive dynamic systems: Cognitive neuroscience and engineering learning from each other,” Proceedings ofthe IEEE, vol. 102, no. 3, pp. 608–628, 2014.