Abstract
Model Predictive Control (MPC) is one of the most successful techniques adopted in industry to control multivariable systems in an optimized way under constraints on input and output variables. In MPC, the manipulated inputs are computed in real-time by solving a mathematical programming problem, most frequently a Quadratic Program (QP). The QP depends on a model of the dynamics of the system, that is often learned from experimental data. To adopt MPC in embedded control systems under fast sampling and limited CPU and memory resources, one must be able to solve QP's with high throughput, using simple code and executing arithmetic operations under limited machine precision, and to provide tight estimates of execution time. Algorithms for solving a convex quadratic program (QP) have been studied since the 1950's, nonetheless it is still a very active subject of research, because of its applications in an extremely large variety of domains, ranging from machine learning to economics, finance, and several engineering fields. In my talk I will present recent advances in embedded quadratic optimization, motivated by (but not limited to) solving MPC problems. I will present and analyze pros and cons of different solution methods, also showing numerical evidence obtained on embedded control hardware platforms under fixed and floating point precision.

Bio
Alberto Bemporad received his master's degree in Electrical Engineering in 1993 and his Ph.D. in Control Engineering in 1997 from the University of Florence, Italy. He spent the 1996/97 academic year at the Center for Robotics and Automation, Department of Systems Science & Mathematics, Washington University, St. Louis, as a visiting researcher. In 1997-1999 he held a postdoctoral position at the Automatic Control Laboratory, ETH Zurich, Switzerland, where he collaborated as a senior researcher in 2000-2002. In 1999-2009 he was with the Department of Information Engineering of the University of Siena, Italy, becoming an associate professor in 2005. In 2010-2011 he was with the Department of Mechanical and Structural Engineering of the University of Trento, Italy. In 2011 he joined the IMT Institute for Advanced Studies Lucca, Italy as a full professor, where he also became the director in 2012. In 2011 he cofounded ODYS S.r.l., a spinoff company of IMT Lucca. He has published more than 250 papers in the areas of model predictive control, hybrid systems, automotive control, multiparametric optimization, computational geometry, robotics, and finance. He is author or coauthor of various MATLAB toolboxes for model predictive control design, including the Model Predictive Control Toolbox (The Mathworks, Inc.). He was an Associate Editor of the IEEE Transactions on Automatic Control during 2001-2004 and Chair of the Technical Committee on Hybrid Systems of the IEEE Control Systems Society in 2002-2010. He received the IFAC High-Impact Paper Award for the 2011-14 triennial. He has been an IEEE Fellow since 2010.

Abstract
Safe Reinforcement Learning (RL) is one of the most widely adopted paradigms to obtain intelligent behavior from interactive robots. RL methods have shown robust and efficient learning on a variety of robot-control problems. However, “the asymptotic nature of guarantees about RL performances makes it difficult to bound the probability of damaging the controlled robot and/or the environment”. How to guarantee that, given a control policy synthesized by RL, such policy will have a very low probability of yielding undesirable behaviors? Our answer leverages Probabilistic Model Checking techniques by describing robot-environment interactions using Markov chains, and the related safety properties using probabilistic logic. Both the encoding of the interaction models and their verification are fully automated, and only the properties have to be manually specified. Our research goes beyond automating verification, to consider the problem of automating repair, i.e., if the policy is found unsatisfactory, it is fixed with no manual intervention.

Abstract
Other the last 15 years nonnegative matrix factorization (NMF) has become a popular unsupervised dictionary learning/adaptive data decomposition technique with applications in many fields. In particular, much research about this topic has been driven by applications in audio, where NMF has been applied with success to automatic music transcription and single channel source source separation. In this setting the nonnegative data is formed by the magnitude or power spectrogram of the sound signal and is decomposed as the product of a dictionary matrix containing elementary spectra representative of the data times an activation matrix which contains the expansion coefficients of the data frames in the dictionary. After a general overview of NMF and a focus on majorization-minimization (MM) algorithms for NMF, the presentation will discuss model selection issues in the audio setting, pertaining to 1) the choice of time- frequency representation (essentially, magnitude or power spectrogram), and 2) the measure of fit used for the computation of the factorization. We will give arguments in support of factorizing of the power spectrogram with the Itakura-Saito (IS) divergence. In particular, IS-NMF is shown to be connected to maximum likelihood estimation of variance parameters in a well- defined statistical model of superimposed Gaussian components and this model is in turn shown to be well suited to audio. Then the presentation will briefly address variants of IS-NMF, namely IS-NMF with regularization of the activation coefficients (Markov model, group sparsity), online IS- NMF, automatic relevance determination for model order selection and multichannel IS-NMF. Audio source separation demos will be played.

Abstract
The voltage-sensitive dye optical imaging (VSDOI) is a promising recording modality for the cortical activity, but its practical potential is limited by many artifacts and interferences in the acquisitions. Inspired by existing models in the literature, we propose a generative model of the signal, based on an additive mixtures of components, each one being constrained within an union of linear spaces, determined by its biophysical origin. Motivated by the resulting component separation problem, which is an underdetermined linear inverse problem, we develop: (1) convex, spatially structured regularizations, extending in particular the popular tools of group sparsity and discrete total variation regularization; (2) a primal, first-order proximal algorithm for minimizing efficiently the resulting functional, which presents both nondifferentiable terms, and terms whose proximity operator cannot be easily computed; (3) statistical methods for automatic parameters selection, based on Stein's unbiased risk estimate. We develop subsequently a software for noisy component separation, and evaluate this software on different VSDOI data set, showing encouraging perspectives for the observation of complex cortical dynamics.

Abstract
I will present an adaptation of the Least Absolute Shrinkage and Selection Operator “LASSO” method to the analysis of correlation dynamics of small neuronal populations. Indeed, due to its low computational cost, Lasso is an attractive regularization method for high dimensional statistical settings. Within our framework, we consider multivariate counting processes depending on an unknown function parameter to be estimated by linear combinations of a fixed dictionary. To select coefficients, we propose an adaptive l1 penalty, where data-driven weights are derived from new Bernstein type inequalities for martingales. Oracle inequalities are established under assumptions on the Gram matrix of the dictionary. This method is then applied to Hawkes processes as model for spike train analysis. The estimation allows us to recover the functional underlying connectivity as the local dependence graph that has been estimated. Simulations and real data analysis show the excellent performances of our method in practice.

Abstract
The recent literature on first order methods for smooth optimization shows that significant improvements on the practical convergence behaviour can be achieved with variable stepsize and scaling for the gradient, making this class of algorithms attractive for a variety of relevant applications. In this paper we introduce a variable metric in the context of the $epsilon$-subgradient projection methods for nonsmooth, constrained, convex problems, in combination with two different stepsize selection strategies. We develop the theoretical convergence analysis of the proposed approach and we also discuss practical implementation issues, as the choice of the scaling matrix. In order to illustrate the effectiveness of the method, we consider a specific problem in the image restoration framework and we numerically evaluate the effects of a variable scaling and of the steplength selection strategy on the convergence behavior.

Bio
Valeria Ruggiero is full Professor in Numerical Analysis at the University of Ferrara. She is Director of the National Group for Scientific Computation of the Istituto Nazionale di Alta Matematica (INdAM) and member of the Scientific Advisory Board of the CINECA Consortium. She was national coordinator of several projects. The research activity of Valeria Ruggiero concerns the development and the analysis of numerical methods for large scale systems, parallel computing, nonlinear optimization and related applications. In particular theoretical and computational results have been obtained about the Inexact Newton interior point method for nonlinear programming problems and nonlinear systems, including the analysis of different iterative solvers for inner linear symmetric indefinite systems. Nonmonotone strategies are analyzed also for the semismooth case, with application to optimal control problems and variational inequalities. More recent research interests are in nonsmooth convex optimization methods for inverse problems, with particular attention to primal-dual first order algorithms for image restoration. The results of her research activity are described in about 60 publications in international scientific journals, proceedings and books, in several software packages and in a number of communications to conferences.

Abstract
This talk considers the recently introduced ordered weighted L1 (OWL) regularizer for sparse estimation problems with correlated variables. We begin by reviewing several convex analysis results concerning the OWL regularizer, namely: that it is indeed a norm, its dual norm, effcient methods to compute the corresponding proximity operator and the Euclidean projection on an OWL ball. We will also show how the OWL norm can be explicitly written as an atomic norm, opening the door to the use of the conditional gradient (Frank-Wolfe) algorithm. In the analysis front, we show that OWL regularization automatically clusters strongly correlated variables, in the sense that the coeffcients associated with such variables have equal estimated values. Furthermore, we characterize the statistical performance of OWL regularization for generative models in which certain clusters of regression variables are strongly (even perfectly) correlated, but variables in different clusters are uncorrelated. We show that if the true p-dimensional signal generating the data involves only s of the clusters, then O(s log p) samples suffce to accurately estimate the signal, regardless of the number of coeffcients within the clusters. The estimation of s-sparse signals with completely independent variables requires just as many measurements. In other words, using the OWL we pay no price (in terms of the number of measurements) for the presence of strongly correlated variables. This work was done in collaboration with Robert Nowak (University of Wisonsin-Madison, USA).

Bio
Mário A. T. Figueiredo received MSc and PhD degrees in electrical and computer engineering,both from Instituto Superior Técnico (IST), the engineering school of the University of Lisbon, in 1990 and 1994. He has been with the faculty of the Department of Electrical and Computer Engineering, IST, since 1994, where he is now a Professor. He is also area coordinator and group leader at Instituto de Telecomunicações, a private non-proft research institute. His research interests include image processing and analysis, machine learning, and optimization. Mário Figueiredo is a Fellow of the IEEE and of the IAPR; he received the 1995 Portuguese IBM Scientifc Prize, the 2008 UTL/Santander-Totta Scientifc Prize, the 2011 IEEE Signal Processing Society Best Paper Award, the 2014 IEEE W. R. G. Baker Award, and several conference best paper awards. His name is included in the Thomson Reuters' Highly Cited Researchers list. He is associate editor of several journals and served as organizer or program committee member of many international conferences.

Abstract
In this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to solve these models efficiently for massive data sets. That challenge is in reach through a new class of variational approximations known as variational compression. The underlying variational bounds are very similar to the objective functions for deep neural networks, giving the promise of efficient approaches to deep learning that are constructed from components with very well understood analytical properties.

Bio
Professor of Machine Learning in the Department of Computer Science at the University of Sheffield and a Professor of Computational Biology in the Sheffield Institute for Translational Neuroscience (part of the Department of Neuroscience).

Abstract
Making accurate predictions is a crucial factor in many systems for cost savings, efficiency, health, safety, and organizational purposes. Taking our motivation from such an increased need in applications, in this talk we present three novel techniques for estimating predictive models from measured data. We start by showing a fully adaptive Tikhonov-type regularization algorithm for addressing the problem of the reconstruction of an unknown functional dependency from given noisy data, primarily focusing on the reconstruction outside the scope of the given data. We provide an efficient recipe of how to choose a regularization space from an admissible set of reproducing kernel Hilbert spaces and a regularization parameter, which allows for an accurate extrapolation from the data. The construction of such an algorithm was motivated by a real-life problem from diabetes therapy management, where the key issue is to predict the future blood glucose levels of a diabetic patient from available current and past information about therapeutically valuable factors. The efficiency and beyond state of the art performance of the algorithm have been demonstrated in extensive numerical experiments with simulated and real clinical data. Inspired by several recent developments in regularization theory, optimization, and image processing, we also present and analyze a new numerical approach to multi-penalty regularization in spaces of sparsely represented functions. We are particularly interested in regularizers, which are able to correctly model and separate the multiple components of additively mixed signals, as they may occur for signals corrupted by additive noise. Finally, we conclude by presenting an ongoing research on a new learning technique to compute inexpensively by very simple operations, without a priori knowledge of the noise level, and with no need of any solution process, a nearly optimal regularization parameter(s) for a given datum of an inverse problem. We exemplify our new technique by showing the learning of nearly optimal soft- thresholding parameters in wavelet shrinkage and TV-based image denoising.

Abstract
The talk will focus on the advantages brought by invariant image representations, especially in the context of deep learning. We will first review image epitomes as a powerful data structure for transformation-aware image analysis. We will then see how we can use epitomes to build image recognition systems invariant to local deformation, both in a bag-of-words (dictionaries of mini-epitomes) and in a deep learning (deep epitomic nets) setting. We will discuss next a practical scheme that incorporates global object position and scale as latent variables into deep epitomic neural networks, both during training and testing. Deep epitomic nets along with explicit scale/position search have been the key ingredients in our TTIC_ECP entry to this year's Imagenet LSVRC image classification competition, achieving 10.2% top-5 error rate, a 3% performance improvement over a baseline conventional max-pooled convnet. (Work done in collaboration in part with L.-C. Chen and A. Yuille (UCLA), and in part with I. Kokkinos and P.-A. Savalle (Ecole Centrale Paris).)

Bio
George Papandreou is a Research Assistant Professor at the Toyota Technological Institute at Chicago (TTI-C), following an appointment (2009-2013) as a postdoctoral research scholar at the University of California, Los Angeles (UCLA). He holds a Diploma (2003) and Ph.D. (2009) in electrical and computer engineering, both from the National Technical University of Athens (NTUA), Greece. His research interests are in computer vision, machine learning, and multimodal perception. In computer vision, he is currently working on invariant patch-based representations for image analysis, labeling, and recognition. In machine learning, he has introduced and is further developing efficient Bayesian inference methods built on top of deterministic energy minimization algorithms. He serves as a reviewer and program committee member to the main journals and conferences in computer vision, image processing, and machine learning and has been a co-organizer of the NIPS 2012 and 2013 Workshops on Perturbations, Optimization, and Statistics.

Abstract
Information theoretic learning is a learning framework that uses descriptors from information theory (entropy and divergence) estimated directly from data to substitute the conventional statistical descriptors of variance and covariance. Minimum error entropy (MEE) is a principle for designing supervised learning algorithms that falls into the information theoretic learning framework. MEE algorithms have been applied successfully in various fields for more than a decade, and can deal with problems involving non-Gaussian noise for which the classical least squares method is not ideal. In this talk we consider empirical MEE learning algorithms in a regression setting. Statistical consistency of an MEE algorithm in an empirical risk minimization framework is presented in details, including error entropy consistency and regression consistency for homoskedastic models and heteroskedastic models. Regularization schemes in reproducing kernel Hilbert spaces are also discussed.

Bio
Ding-Xuan Zhou is a Chair Professor and Head of Department of Mathematics at City University of Hong Kong, Hong Kong. He received the B.S. and Ph.D. degrees in applied mathematics from Zhejiang University, China, in 1988 and 1991, respectively. In 1992, he was a Postdoctoral Fellow at the Institute of Mathematics, Chinese Academy of Sciences in Beijing. He was an Alexander von Humboldt Research Fellow at University of Duisburg, Germany from 1993 to 1995, and a Postdoctoral Fellow and Instructor at University of Alberta, Canada, from 1995 to 1996. He is mainly interested in learning theory, wavelet analysis, and approximation theory. In 2005, he received a Joint Research Fund for Hong Kong and Macau Young Scholars from the National Science Fund for Distinguished Young Scholars of the National Science Foundation of China. He has received numerous research grants from the Research Grants Council of Hong Kong and organized many international conferences. He serves the editorial boards of the journals: Advances in Computational Mathematics, Applied and Computational Harmonic Analysis, Analysis and Applications, Complex Analysis and Operator Theory, Journal of Computational Analysis and Applications, and Oriental Journal of Mathematics.

Abstract
Optimal transport distances are a fundamental family of distances for histograms and points clouds. Despite their appealing theoretical properties, excellent performance in retrieval tasks and intuitive formulation, their computation involves the resolution of a linear program whose cost is prohibitive whenever the histograms' dimension exceeds a few hundreds. We propose in this work a new family of optimal transport distances that look at transport problems from a maximum-entropy perspective. We smooth the classical optimal transport problem with an entropic regularization term, and show that the resulting optimum is also a distance which can be computed through Sinkhorn-Knopp's matrix scaling algorithm at a speed that is several orders of magnitude faster than that of transport solvers. We show how and why this approach can be effective regardless of the properties of the ground metric, and how it can be directly applied to provide new computational approaches to solve variational problems that involve optimal transport distances, such as the Wasserstein barycenter problem recently proposed by Agueh and Carlier (2010).

Bio
Marco Cuturi received his Ph.D. in applied maths from the Ecole des Mines de Paris under the supervision ofJean-Philippe Vert. After working at the ORFE department of Princeton University between 02/2009 and 08/2010 as a lecturer, he joined the Graduate School of Informatics in 09/2010 as a G30 associate professor. He is currently the associate professor of the Yamamoto-Cuturi lab, starting from 11/2013.

Abstract
Statisticians in the 21st century can look forward to very bright futures for the discipline and profession of statistics by solving the important problem of teaching thousands statistical data scientist aspirants, what are the fundamental methods of statistical learning. We are developing a framework that unifies parametric and nonparametric models, frequentist and Bayesian inference, small data and big data, beginning courses that differ from traditional courses by teaching methods for simple data in ways that extend to complex high dimensional data (similar to the goal of teaching finite dimensional math in the notation that extend to Hilbert space). To accomplish the above, a theory, called LP STATISTICAL SCIENCE, is being developed by myself and Subhadeep (Deep) Mukhopadhyay (Temple University Fox School of Business). The theory is unifying because it applies to mixed data (variables which are discrete or continuous) and to population and sample distributions. It practices 'plug-in estimation' computed for sample distributions by the same definitions used for population distributions.

Bio
Emanuel Parzen is a Distinguished Professor Emeritus, Dept. of Statistics, Texas A&M University. He has served as an Assistant Professor of Mathematical Statistics at Columbia, Professor of Statistics at Stanford, a Leading Professor-Chairman of the Dept. of Statistics and a Director of Statistical Science at the State University of New York at Buffalo, and a Distinguished Professor at Texas A&M University. He has been a Fellow at Imperial College London, IBM Systems Research Institute, the Center for Advanced Study in the Behavioral Sciences at Stanford, and a Visiting Professor at the Sloan School of MIT, the Dept. of Statistics at Harvard and the Dept. of Biostatistics at Harvard. Professor Parzen has authored or coauthored over 120 papers and 6 books, on signal detection theory, time series analysis and non-parametric statistics. He pioneered the use of kernel density estimation (also known as the Parzen window in his honor) and Reproducing Kernel Hilbert Spaces (RKHS) in statistics and engineering. He has served on innumerable editorial boards and national committees, and has organized several influential conferences and workshops. He has been elected Fellow of the American Statistical Association, the Institute of Mathematical Statistics and the American Association for the Advancement of Science. In 1994, he was awarded the prestigious Samuel S. Wilks Memorial Medal by the American Statistical Association.

Abstract
Collective intelligence systems, from prediction markets to Wikipedia, have the capacity to provide useful information by aggregating the wisdom of the crowd. Yet the mechanisms that govern how individuals interact in these forums can substantially affect the quality of information produced. I will discuss this issue in the context of two specific problems in prediction markets: ensuring sufficient liquidity and mitigating manipulation. The accuracy of the information reflected in market prices depends on the market’s liquidity. In a liquid market, arriving traders have someone to trade with at a 'reasonable' price, so they are willing to participate and contribute their information. Liquidity provision can be framed as a reinforcement learning problem for a market-making agent, complicated by the censored nature of observations. I will describe an algorithm for solving this problem using moment-matching approximations in the belief space, and discuss theoretical results and empirical evaluation of the algorithm in experiments with trading agents and human subjects, showing that it offers several potential benefits over standard cost-function based approaches. In markets where participants influence the outcome of the events on which they are trading, concerns over manipulation naturally arise. I will present a game-theoretic model of manipulation, which gives insight into the question of how informative market prices are in the presence of manipulation opportunities, and also into how markets can affect the incentives of agents in the outside world. In addition, I will describe our experience with a field experiment related to manipulation, the Instructor Rating Markets. Time permitting, I will also briefly discuss work in my group on related issues in other types of collective intelligence systems, for example, information growth, user engagement, and manipulation in social media like Wikipedia and Reddit.

Bio
Sanmay Das is an Associate Professor with the Department of Computer Science and Engineering, Washington University in St. Louis. He was previously an Assistant Professor at Rensselaer Polytechnic Institute and an Associate Professor at Virginia Tech. He is the recipient of an NSF CAREER Award and is currently vice-chair of the ACM Special Interest Group on Artificial Intelligence. He has served as program co-chair of AMMA, as workshop chair of ACM EC, and as sponsorship co-chair of AAMAS, in addition to serving on the program committees and senior program committees of many conferences in artificial intelligence and machine learning, including AAAI, IJCAI and ICML. His research lies in designing effective algorithms for agents in complex, uncertain environments and in understanding the social or collective outcomes of individual behavior. He has broad interests across computational social science (market microstructure, matching markets, social networks) and machine learning (reinforcement learning, sequential decision-making, supervised learning, data mining).

Abstract
Behaviors, from simple to most complex, require a two-way interaction with the environment and the contribution of different brain areas depending on the orchestrated activation of neuronal assemblies. Understanding how information is coded and how synaptic mechanisms are implemented in these networks is one of the major challenges of neuroscience. Our research is focused on in vitro systems, constituted by neuronal cultures coupled to Micro-Electrode Arrays (MEAs). The goal is to characterize the electrophysiological behavior of large neuronal assemblies with the aim of understanding how these ‘reduced’ networks develop, learn and modify the strength of their synapses. In the first part of the talk, V. Pasquale will present results of experiments aimed at elucidating the mechanisms of generation and propagation of spontaneous activity and the interplay between spontaneous and evoked events within neuronal networks. Recent studies on that topic reported the existence of privileged neurons that consistently fire earlier than others at the onset of synchronized bursting events (or network bursts, NB), which have been termed major burst leaders (MBL). At the same time, by stimulating the network from different channels one can obtain very different responses, not only in size and delay but also in the activation order of the responding neurons. Thanks to her experiments, she will demonstrate that MBLs do not only drive the propagation of coordinated spontaneous activations, but also play a special role in coordinating and driving the evoked bursts of activity.In the second part of the talk, M. Chiappalone will introduce the main research topics in which she has been involved from her PhD studies to the current studies in the lab settled at the Italian Institute of Technology. Specifically, she will introduce the concept of ‘embodied electrophysiology’, an interdisciplinary experimental framework at the interface between neuroscience and robotics, which has the primary goal of understanding the mechanisms of adaptive behavior and neural coding. She will explain how to take advantage of in vitro systems constituted by primary neuronal cultures coupled to Micro-Electrode Arrays (MEAs) for studying closed loop-interaction. More specifically, in the lab a closed loop system involving a neuronal culture and a small robot has been developed. Such an experimental framework is well suited for investigations on neural coding and in the Neural Interface (NI) field of research. She will then discuss the possibility to use a similar closed-loop experimental paradigm for developing a ‘brain-prosthesis’, aimed at restoring lost functions in different experimental models at increasing anatomical complexity. This is also the main objective of the recently funded FET Young Explorers EU project ‘Brain Bow’ (http://www.brainbowproject.eu/), which she is coordinating.

Bio
Michela Chiappalone graduated in Electronic Engineering (summa cum laude) at the University of Genova (Italy) in 1999 and obtained a PhD in Electronic Engineering and Computer Science from the same University in 2003. During the last year of her PhD training, she has been visiting scholar at the Dept. of Physiology of the Northwestern University (Chicago, IL, USA), under the supervision of Prof F.A. Mussa-Ivaldi, working on a neuro-robotic system involving the brain of a sea lamprey connected to a small mobile robot. From 2003 to 2006 she got a fellowship from the University of Genova under the supervision of Prof. S. Martinoia, being involved in several national and international projects on neuroengineering studies. In 2004 she won the Italian “PATRON PhD Award” for the best PhD dissertation in the bioengineering field.From 2007 to 2012 she has been Post-Doc at the Italian Institute of Technology (IIT) in the Neuroscience and Brain Technologies Dept. (NBT), led by Prof. F. Benfenati and Prof. J. Assad. At the present time, she holds a ‘Researcher’ position and leads the ‘Neural Interface and Network Electrophysiology’ (NINE) lab in the same Institution. In 2012 M. Chiappalone has been awarded a FET Young Explorers grant (‘BRAIN BOW’, www. bainbowproject.eu) by the European Commission. Her main research interests include: (1) Studying dynamics and plasticity of in vitro neuronal assemblies; (2) Using in vitro cultures as a tool for high-throughput neuropharmacological and neurotoxicological screening (3) Establishing a bi-directional communication between natural and artificial elements (4) Developing algorithms for multichannel data analysis. M. Chiappalone contributed to all these activities with 40 publications in International Journals, 5 book chapters and more than 70 contributions at international conferences.

Abstract
Examples of application of machine learning techniques will be given in different fields of Control Theory. In particular, the following topics will be addressed: - Distributed information control problems. - Moving horizon state estimation. - Control strategies for the maximization of Information gain - Open problems in mode identification of switching systems.

Abstract
Already early in development, infants learn to solve visual problems that are highly challenging for even the most sophisticated computational methods. Two striking examples are learning to recognize hands and learning to follow other people's gaze direction. I will present a model that can imitate infants learning of these concepts. It is shown a stream of natural videos and it learns without any supervision to detect human hands as well as direction of gaze, in complex natural scenes. The key to the successful learning comes from the use by the system of simple innate biases, which guide the learning process along a path that leads it to acquire useful representations and sophisticated detection processes. I will discuss some general implications to the interactions between learning and innate mechanisms, and the use of innate cognitive mechanisms to make artificial systems better learners. This work is a part of a broader project called "digital baby" which will be briefly described.

Bio
Shimon Ullman is the Ruth and Samy Cohn Professor of Computer Science in the department of computer science and applied mathematics at the Weizmann Institute of Science in Israel. He received his Bs.C. from the Hebrew University in Jerusalem, and Ph.D. from M.I.T, where he has been a Professor in the Brain and Cognitive Science Department and the Artificial Intelligence Laboratory. His main areas of research are human vision and cognition, brain modeling, and computer vision. He is the 2008 recipient of the Rumelhart award and a member of Israeli Academy of Science.

Abstract
The recent developments in robot sensors and actuators bring a new human-centric perspective to robotics. It offers new applications in which the robots must work in safe collaboration with the users, by generating natural movements and anticipating their behaviors. The variety of signals to process and the richness of interaction with the users and the environment constitute a formidable area of research for machine learning. Current programming solutions used by the leading commercial robotics companies do not satisfy the requirements of re-using the same robot for different tasks. The representation of manipulation movements must be augmented with forces (for task execution, but also as a communication channel for collaborative manipulation), compliance and reactive behaviors. An attractive approach to the problem of transferring skills to robots is to take inspiration from the way humans learn by imitation and self-refinement. I will present an approach based on a task-parameterized Gaussian mixture model, which is combined with optimal control and stochastic optimization for reproduction and refinement of an observed skill. The representation as a mixture model facilitates the combination and reorganization of the demonstrated tasks in different ways. The extension to a task-parameterized model allows the robot to generalize the learned task to a new situation, by observing the same set of demonstrations from the perspective of multiple coordinate systems (typically, different position/orientation of tools or objects considered as candidate frames). Finally, the probabilistic representation allows us to take into account multiple observations of the same task, and exploit this information within an optimal control strategy. The result is a minimal intervention controller regulating the stiffness and damping of the robot according to the precision requirements that can change during the task. Namely, the robot will be controlled stiffly only in the directions and when it requires to be accurate, and will otherwise remain compliant (safer for the end-user and energy efficient). Examples of applications with a compliant humanoid and with gravity-compensated manipulators will be showcased.

Bio
Dr Sylvain Calinon is Team Leader of the Learning & Interaction Lab at the Department of Advanced Robotics at IIT since 2009. He holds a PhD from the Ecole Polytechnique Federale de Lausanne (2007) awarded by the Robotdalen Scientific Award, ABB Award and EPFL-Press Distinction. From 2007 to 2009, he was Postdoctoral Fellow at the Learning Algorithms and Systems Lab, EPFL. He has published 60 publications and a book in the field of robot learning by imitation and human-robot interaction.

Abstract
Many phenomena observed in speech, such as, e.g., coarticulation effects, can be easily and compactly described in terms of vocal tract gestures but not in purely acoustic terms. That has been a strong motivation to use knowledge of the behavior of the vocal tract (when it produces speech) for Automatic Speech Recognition (ASR). The approaches using measured articulatory data of the vocal tract typically require an acoustic-to-articulatory mapping (AAM) to be learned in order to recover the articulatory information from speech acoustics (since only acoustics is available during recognition). In this talk I will present our work on AAM for phone recognition where AAM processes go through multi-layered and hierarchical (i.e., deep) representations of the acoustic and the articulatory domains. Cross-corpus and cross-linguistic results show that recovered articulatory features consistently increase recognition performance in a hybrid speaker-dependent Deep Neural Network � Hidden Markov Model phone recognition system.