Abstract
For functions chosen from a union of classes there is a trick using concentration inequalities to improve generalization bounds. The advantages of the method are particularly pronounced in high dimensions. I will explain the trick and apply it to examples of learning with structured sparsity and some versions of multitask learning.

Abstract
In this talk, I will present a general vector-valued reproducing kernel Hilbert spaces (RKHS) framework for the problem of learning an unknown functional dependency between a structured input space and a structured output space. Our formulation encompasses both Vector-valued Manifold Regularization and Co-regularized Multi-view Learning, providing in particular a unifying framework linking these two important learning approaches. In the case of the least square loss function, we provide a closed form solution, which is obtained by solving a system of linear equations. In the case of Support Vector Machine (SVM) classification, our formulation generalizes in particular both the binary Laplacian SVM to the multi-class, multi-view settings and the multi-class Simplex Cone SVM to the semi-supervised, multi-view settings. Along with the mathematical formulations, I will present empirical results obtained on the task of object recognition using several challenging datasets.

Bio
Ha Quang Minh received the BSc degree from Monash University (Melbourne, Australia) and MSc and Ph.D. degrees in mathematics from Brown University, Providence, RI, USA. He is currently a Senior Post-Doctoral Researcher in pattern analysis and computer vision at the Istituto Italiano di Tecnologia (IIT), Genova, Italy. Prior to joining IIT, he was at the University of Chicago, University of Vienna (Austria), and Humboldt University of Berlin (Germany). His current research interests include applied and computational functional analysis, machine learning, and applications in data analysis, computer vision, image and signal processing.

Abstract
In this talk I shall review some recent work on the use of Bayesian/regularization techniques for estimation of dynamical systems. The distinctive features of dynamical systems which should be accounted for will be discussed. Furthermore some recent results and insights on the design and data based tuning of priors/kernels will be given discussing the link with SURE-type estimators. Future challenges will be illustrated.

Bio
Associate Professor with the Dipartimento di Tecnica e Gestione dei Sistemi Industriali at the Universita` di Padova. His research interests are mainly in Estimation, Identification Theory and Applications (subspace methods, stochastic realization, non-linear estimation, hybrid systems, adaptive optics), Computer Vision (structure from motion, texture and gait analysis) and Networked Estimation and Control.

Bio
Massimiliano Pontil received an MSc degree in Physics from the University of Genova in 1994 (summa cum laude) and a PhD in Physics from the same University in 1999. He spent approximately half of the PhD studies at the Massachusetts Institute of Technology (MIT) as a Visiting Researcher. Massimiliano is Professor and EPSRC Advanced Research Fellow in the Department of Computer Science at University College London (UCL). At UCL he has also been a Lecturer, between January 2003 and September 2006, and a Reader between October 2006 and September 2010. Before joining UCL, Massimiliano was a Research Associate in the Department of Information Engineering at University of Siena (2001--2002) and a Post-doctoral Fellow in the Center for Biological and Computational Learning at the Massachusetts Institute of Technology (MIT) (1998--2000). He has also been a Visiting Fellow at the Isaac Newton Institute for Mathematical Sciences in Cambridge, at the Catholic University of Leuven, at the University of Chicago and at the City University of Hong Kong, among others. His research interests are in the area of machine learning and pattern recognition, with a focus on regularization methods, convex optimization and statistical estimation. He also studied machine learning applications arising in Computational Vision, Natural Language Processing and Bioinformatics. He has published about hundred papers in the above research areas, has been on the programme committee of the main machine learning conferences, including COLT (2005, 2006, 2008, 2009, 2010) and ICML (2004, 2009) and is an Associate Editor of the Machine Learning Journal, Statistics and Computing and Action Editor for the Journal of Machine Learning Research.

Abstract
High dimensional data appears in a wide variety of applications, from signal processing (e.g. sounds, images), to the study of dynamical systems with high dimensional state spaces, to the study of corpora of text documents, to medical, biological and financial data. A basic model is to think of a data point as a sample from a high-dimensional probability distribution. Traditional statistical methods fail in high-dimensions due to the curse of dimensionality, and new hypotheses on the structure of data are needed. In particular, the assumption that data, while presented in high-dimensions, is intrinsically low-dimensional, has been verified in many data sets, and has been useful in deriving new methods in statistics and machine learning. We will discuss techniques that analyze the geometry of data in a robust manner (both with respect to sample size and high-dimensional noise) to estimate the intrinsic dimension, efficiently construct representations of data and dictionaries for it, and to estimate the probability distribution generating the data. We present applications to the detection of anomalies in hyper-spectral images, images, and in regression and classification problems.

Bio
Professor in Mathematics, Electrical and Computer Engineering, Computer Science, Duke University. His recent work has focused on the construction of multi-resolution structures on discrete data and graphs, connecting aspects of classical harmonic analysis, global analysis on manifolds, spectral graph theory and classical multiscale analysis

Abstract
Algorithms are exposed to randomness in the input or noise during the computation. How well can they preserve the information in the data w.r.t. the output space? Algorithms especially in Machine Learning are required to generalize over input fluctuations or randomization during execution. This talk elaborates a new framework to measure the "informativeness" of algorithmic procedures and their "stability" against noise. An algorithm is considered to be a noisy channel which is characterized by a generalization capacity (GC). The generalization capacity objectively ranks different algorithms for the same data processing task based on the bit rate of their respective capacities. The problem of grouping data is used to demonstrate this validation principle for clustering algorithms, e.g. k-means, pairwise clustering, normalized cut, adaptive ratio cut and dominant set clustering. Our new validation approach selects the most informative clustering algorithm, which filters out the maximal number of stable, task-related bits relative to the underlying hypothesis class. The concept also enables us to measure how many bit are extracted by sorting algorithms when the input and thereby the pairwise comparisons are subject to fluctuations.

Bio
Joachim M. Buhmann leads the Machine Learning Laboratory in the Department of Computer Science at ETH Zurich. He has been a full professor of Information Science and Engineering since October 2003. He studied physics at the Technical University Munich and obtained his PhD in Theoretical Physics. As postdoc and research assistant professor, he spent 1988-92 at the University of Southern California, Los Angeles, and the Lawrence Livermore National Laboratory. He held a professorship for applied computer science at the University of Bonn, Germany from 1992 to 2003. His research interests spans the areas of pattern recognition and data analysis, including machine learning, statistical learning theory and information theory. Application areas of his research include image analysis, medical imaging, acoustic processing and bioinformatics. Currently, he serves as president of the German Pattern Recognition Society.

Abstract
Today�s AI technologies, such as Watson and Siri, are impressive yet still confined to a single domain or task. Imagine how truly intelligent systems---systems that actually understand their world---could change our world. A successful research plan for understanding intelligence includes two key domains: the domain of the physical world and the domain of human agents and their interactions. First, understanding intelligence requires scene, object and action recognition; second, it requires non-verbal social perception (NVSP). As an example of research in the first domain, I will describe work at the joint IIT-MIT Laboratory for Computational and Statistical Learning over the last 2 years developing a theory of visual cortex and of deep learning architectures of the convolutional type. I will describe the theoretical consequences of a simple assumption: the main computational goal of the feedforward path in the ventral stream � from V1, V2, V4 and to IT � is to discount image transformations, after learning them during development. A basic neural operation consists of dot products between input vectors and synaptic weights � which can be modified by learning. I will outline theorems showing that a multi-layer hierarchical architecture of dot-product modules can learn in an unsupervised way geometric transformations of images and then achieve the dual goals of invariance to global affine transformations and of robustness to deformations. These architectures can achieve invariance to transformations of a new object: from the point of view of machine learning they show how to learn in an unsupervised way representations that may reduce considerably the sample complexity of supervised learning.

Bio
Tomaso Poggio is the Eugene McDermott Professor at the Department of Brain and Cognitive Sciences; Co- Director, Center for Biological and Computational Learning; Member of the Computer Science and Artificial Intelligence Laboratory at MIT; He is one of the founders of computational neuroscience. He pioneered models of the fly�s visual system and of human stereovision, introduced regularization theory to computational vision, made key contributions to the biophysics of computation and to learning theory, developed an influential model of recognition in the visual cortex.

Abstract
In this talk I will discuss the problem of estimating the ratio q(x)/p(x), where p and q are density functions given by sampled data. This ratio appears in a number of different settings, including the classical importance sampling in statistics and, more recently, in transfer learning, where an inference procedure learned in one domain needs to be generalized to other tasks. Our method is based on posing the ratio estimation as an inverse problem expressed by a Fredholm integral equation. This allows us to apply the classical techniques of regularization and to obtain simple and easy implementable algorithms within the kernel methods framework. We provide detailed theoretical analysis for the case of the Gaussian kernel and show very competitive experimental comparisons in several settings.

Bio
Mikhail Belkin is an Associate Professor at the Computer Science and Engineering Department and the Department of Statistics at the Ohio State University. His research focuses on applications and theory of machine and human learning. He received his Ph.D degree from the Mathematics Dept. at University of Chicago in 2003. He received the U.S National Science Foundation (NSF) Career Award in 2007, and the Lumley Research Award at the College of Engineering of OSU in 2011. He is on the Editorial board for Journal of Machine Learning Research (JMLR) and IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). Currently he is on sabbatical at the Austrian Institute of Technology (ISTA).

Abstract
Intelligent autonomous robots that can assist humans in situations of daily life have been a long standing vision of robotics, artificial intelligence, and cognitive sciences. A elementary step towards this goal is to create robots that can learn tasks triggered by environmental context or higher level instruction. However, learning techniques have yet to live up to this promise as only few methods manage to scale to high-dimensional manipulator or humanoid robots. In this talk, we investigate a general framework suitable for learning motor skills in robotics which is based on the principles behind many analytical robotics approaches. It involves generating a representation of motor skills by parameterized motor primitive policies acting as building blocks of movement generation, and a learned task execution module that transforms these movements into motor commands. We discuss learning on three different levels of abstraction, i.e., learning for accurate control is needed to execute, learning of motor primitives is needed to acquire simple movements, and learning of the task-dependent "hyperparameters" of these motor primitives allows learning complex tasks. We discuss task-appropriate learning approaches for imitation learning, model learning and reinforcement learning for robots with many degrees of freedom. Empirical evaluations on a several robot systems illustrate the effectiveness and applicability to learning control on an anthropomorphic robot arm. A large number of real-robot examples will be demonstrated ranging from Learning of Ball-Paddling, Ball-In-A-Cup, Darts, Table Tennis to Grasping.

Bio
Jan Peters is electrical and mechanical engineer (Dipl-Ing, TU Muenchen; MSc USC) and computer scientist (Dipl-Inform, FernUni Hagen; MSc, PhD USC) who was educated and performed research at the TU Muenchen, the DLR Robotics Center in Germany, at ATR in Japan, at USC in California. Between 2007 and 2010, he was a senior research scientist and group leader at the Max-Planck Institute for Biological Cybernetics. Since 2011, he is a senior research scientist and group leader at the Max-Planck Institute for Intelligent Systems and a full professor at Technische Universitat Darmstadt.