Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref A PPLICATIONS IN N EUROSCIENCE , I NDUSTRY 4.0, AND S PORTS Pedro Larra˜ naga Computational Intelligence Group Artificial Intelligence Department Universidad Polit´ ecnica de Madrid Bayesian Networks: From Theory to Practice International Black Sea University Autumn School on Machine Learning 3-11 October 2019, Tbilisi, Georgia Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 1 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Outline 1 Introduction 2 Bayesian Networks 3 Neuroscience 4 Industry 5 Sport 6 Estimation of Distribution Algorithms 7 Conclusions 8 References Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 2 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Outline 1 Introduction 2 Bayesian Networks 3 Neuroscience 4 Industry 5 Sport 6 Estimation of Distribution Algorithms 7 Conclusions 8 References Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 3 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Machine learning CONNECTIONISTS Backpropagation EVOLUTIONARIES Genetic programming SYMBOLISTS Inverse deduction BAYESIANS Probabilistic inference ANALOGIZERS Kernel machines The five tribes of machine learning and their master algorithms (Domingos, 2015) Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 4 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Machine learning Multi-output regression Clustering Multiview-clustering • Problem transformation methods • Agglomerative clustering • Non-probabilistic – Single-target method – Hierarchical • Density-based – Multi-target regressor stacking • Agglomerative • Principal component analysis – Regressor chains • Divisive – Canonical correlation analysis – Partitional • Algorithm adaptation methods • Spectral clustering • K -means – Multi-output suppor vector regression • K -medians • Co-regularization – Kernel methods • K -modes • Ensemble clustering – Multi-target regression trees • Fuzzy C-means – Rule method • Self-organizing map • Bayesian networks • Spectral clustering – Gaussian Bayesian networks • K -medoids • Affinity propagation Feature subset selection • K -plane clustering • Fuzzy C-shell • Filter approaches • DBSCAN – Univariate filters • Probabilistic – Multivariate filters Supervised classification – Finite-mixture models • Wrapper methods • Non-probabilistic – Bayesian networks – Bayesian networks-based EDAs – Nearest neighbors • Embedded methods – Classification trees – Rule induction MACHINE • Hybrid methods – Artificial neural networks LEARNING – Support vector machines Discovery associations • Probabilistic • Association rules – Discriminant analysis • Bayesian networks Multi-dimensional classifiers – Logistic regression – Bayesian network classifiers • Problem transformation methods Anomaly detection Anomaly detection • Metaclassifiers – Binary relevance • Probabilistic approaches • Probabilistic approaches – Fusion of outputs – Classifier chains – Stacked generalization • Distance-based • Distance-based – RA k EL – Cascading – LPBR • Reconstruction-based • Reconstruction-based – Label powerset – Bagging • Domain-based • Domain-based – Random forest • Algorithm adaptation methods • Information theory-based • Information theory-based – Boosting • Multi-dimensional Bayesian network clas- – Hybrid classifiers • Bayesian network likelihood • Bayesian network likelihood sifiers Machine learning methods for approaching the eight problems in this talk Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 5 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Machine learning Multi-output regression Clustering Multiview-clustering • Problem transformation methods • Agglomerative clustering • Non-probabilistic – Single-target method – Hierarchical • Density-based – Multi-target regressor stacking • Agglomerative • Principal component analysis – Regressor chains • Divisive – Canonical correlation analysis – Partitional • Algorithm adaptation methods • Spectral clustering • K -means – Multi-output suppor vector regression • K -medians • Co-regularization – Kernel methods • K -modes • Ensemble clustering – Multi-target regression trees • Fuzzy C-means – Rule method • Self-organizing map • Bayesian networks • Spectral clustering – Gaussian Bayesian networks • K -medoids • Affinity propagation Feature subset selection • K -plane clustering • Fuzzy C-shell • Filter approaches • DBSCAN – Univariate filters • Probabilistic Supervised classification – Multivariate filters – Finite-mixture models • Wrapper methods • Non-probabilistic – Bayesian networks – Bayesian networks-based EDAs – Nearest neighbors – Classification trees • Embedded methods – Rule induction MACHINE • Hybrid methods – Artificial neural networks – Support vector machines LEARNING Discovery associations • Probabilistic • Association rules – Discriminant analysis – Logistic regression • Bayesian networks Multi-dimensional classifiers – Bayesian network classifiers • Problem transformation methods Anomaly detection Anomaly detection • Metaclassifiers – Binary relevance • Probabilistic approaches • Probabilistic approaches – Fusion of outputs – Classifier chains – Stacked generalization • Distance-based • Distance-based – RA k EL – Cascading – LPBR • Reconstruction-based • Reconstruction-based – Label powerset – Bagging • Domain-based • Domain-based – Random forest • Algorithm adaptation methods • Information theory-based • Information theory-based – Boosting • Multi-dimensional Bayesian network – Hybrid classifiers • Bayesian network likelihood • Bayesian network likelihood classifiers Machine learning methods for approaching the eight problems in this talk Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 6 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref The 23 Asilomar AI principles + Explainable Artificial Intelligence Asilomar Conference on Beneficial AI (Future of Life Institute 2017) 1. Research goal 2. Research funding 3. Science-policy lik 4. Research culture 5. Race avoidance 6. Safety 7. Failure transparency 8. Judicial transparency 9. Responsability 10. Value alignment 11. Human values 12. Personal pricavy 13. Liberty and privacy 14. Shared Benefit 15. Shared prosperity 16. Human control 17. Non-subversion 18. AI arms race 19. Capacity caution 20. Importance 21. Risks 22. Recursive self-improvement 23. Common good Explainable Artificial Intelligence (Rudin, 2019) Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 7 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Outline 1 Introduction 2 Bayesian Networks 3 Neuroscience 4 Industry 5 Sport 6 Estimation of Distribution Algorithms 7 Conclusions 8 References Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 8 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref “Risk of dementia” (Bielza and Larra˜ naga, 2014a) Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 9 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref “Risk of dementia” p ( A , N , S , D , P ) = p ( A ) p ( N | A ) p ( S | A ) p ( D | N , S ) p ( P | S ) Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 10 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref “Risk of dementia” p ( A , N , S , D , P ) = p ( A ) p ( N | A ) p ( S | A ) p ( D | N , S ) p ( P | S ) Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 10 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Inference (reasoning) with Bayesian networks No evidence Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 11 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Inference (reasoning) with Bayesian networks Evidence: “Stroke = yes” Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 12 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Inference (reasoning) with Bayesian networks Evidence: “‘Stroke = yes, Neuronal Atrophy=yes” Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 13 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Inference (reasoning) with Bayesian networks Evidence: “‘Stroke = yes, Neuronal Atrophy=yes, Age= young” Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 14 / 88
Intro Bayes Nets Neuro Industry Sport EDAs Concl Ref Bayesian networks (Pearl, 1988; Koller and Friedman, 2009) Structure and parameters A Bayesian network consists of two components 1 Graphical structure G is a directed acyclic graph (DAG) Vertices → variables Directed edges → conditional (in)dependences 2 Set of parameters specifies the set of conditional probability distributions Joint probability distribution: P ( x 1 , . . . , x n ) = � n � x i | pa ( x i ) � i = 1 P Pedro Larra˜ naga Neuroscience, Industry 4.0, and Sports 15 / 88
Recommend
More recommend
