Lecture 1: Latent Variable Models : One of central problems in machine learning is that of density estimation, which involves constructing a model of a probability distribution, given a finite number of samples drawn from that distribution. A latent variable model is a powerful approach to probabilistic modeling, supplementing a set of observed (visible) variables with additional latent (hidden) variables. In this tutorial, I will introduce widely-used latent variable models, including linear generative models and mixture models. For linear generative models, I will first explain factor analysis where the goal is to seek a linear generative model with Gaussian factors which best models the covariance structure of data variables. Independent component analysis is a powerful spin-off of factor analysis, allowing non-Gaussian factors in the model. I will also introduce an important family of latent variable models, that is known as mixture models, including mixture of Gaussians, mixture of factor analyzers, and mixture of experts. All these aforementioned latent variables models are parametric models where parameters are learned by maximum likelihood estimation. Expectation maximization (EM) is a powerful technique for learning parameters in the presence of latent variables, which will be also discussed. This tutorial focuses on basic concept and fundamental mathematical theory involving latent variable models and EM optimization, rather than any realistic applications, in order to give audience in robotics a taste of machine learning, stressing out that machine learning will be the best friend of robotics in near future.

Lecture 2: Support Vector Machine : With a great deal of recent theoretical and empirical successes, the support vector learning method has grown up as a viable tool in the area of intelligent systems. Among the important application areas for the support vector learning, we have SVC(support vector classification), SVR(support vector regression), and SVDD(support vector data description). In this tutorial, we will present fundamentals on SVC, SVR, and SVDD, together with related topics such as kernel PCA(principal component analysis) and RKHS(reproducing kernel Hilbert space) theory. After briefly presenting the above fundamentals, we will also introduce some recent results on the pre-image problems and the kernel-based de-noising methods.

Lecture 3: Bayesian Netowkrs : Bayesian networks are a kind of probabilistic graphical models that compactly represent the joint probability distribution over a set of random variables via a comprehensible DAG (directed-acyclic graph) structure. A Bayesian network can be applied to both prediction and description tasks. Thus, Bayesian networks have been widely applied in various areas, spanning from robotics and speech recognition to medical diagnosis and bioinformatics. This tutorial will cover the basic building blocks for deploying Bayesian networks in real-world applications. In specific, various approaches to learning Bayesian networks from data are introduced. As an example application, DNA microarray data analysis using Bayesian networks is presented. In addition, a couple of advanced learning techniques for alleviating the small-sample problem are discussed.