Guia docente 2013_14 Escola de Enxeñaría de Telecomunicación

Máster Universitario en Teoría do Sinal e Comunicacións.

Recoñecemento Estatístico de Patróns e Redes Neuronais

Contents

Topic	Sub-topic
Introduction	Approximation to the problem of pattern recognition. Review of Probability Theory and Rule of *Bayes
Classical concepts of classification and dimensionality reduction	Unsupervised classification or clustering. K-means algorithm. Non parametric supervised classification. K nearest neighbour algorithm. Statistical classification. Minimum distance classifier. Optimum Bayes classifier. Methods of extraction of characteristics: optimisation for representation (PCA), optimisation for classification (LDA)
Gaussian Mixture Models to estimate probability density functions	Gaussian Mixture Models for representation and for classification. Esimate of maximum likelihood for the model: The EM algorithm. Particular cases. Application to speech and speaker recognition: Hidden Markov Models.
Learning processes and introduction to the artificial neural networks	Foundamentals of learning theory. The statistical nature of the learning process. Usual Learning Rules. Concepts of Learning theory: approximation error, estimation error and calculation error. Bias and variance of models. Llearning techniques: correction error, Hebb rule, competition and supervision. Taxonomy of ANN. Discriminative Models versus Generative Models.
The multilayer perceptron (MLP).	The perceptron rule. Theorem of convergence. Separability, the XOR problem. Minimisation of the Mean Square error. The multilayer perceptron. The backpropagation algorithm. The generalisation problem, cross-validation. Interpretation of the outputs as a posteriori probabilities.
Radial Basis Function (RBF).	Cover´s theorem on the separability of patterns. The interpolation problem. Regularisation theory. Generalized Radial Basis Functions. Strategies of learning. Comparison between RBF and MLP. Analogy between RBF-GMM (Discrimination versus representation)
Support Vector Machines (SVM).	Classifiers of maximum margin. The dimension of Vapnik-Chervonenkis. Kernel-based spaces of characteristics. SVM for binary classification (SVC). SVM for non-linear regression (SVR). SVM for clustering (SVND).
Self-organized Networks.	Hebbian learning network: analysis of principal components. Maps of self-organized features, adaptative learning classifiers, Learning vector quantization (LVQ). Autoassociative Networks.