Inspired by the success of compressive sensing, the last three years have seen an explosion of research in the theory of low-rank modeling. By now, we have results stating that it is possible to recover certain low-rank matrices from a minimal number of entries -- or of linear functionals -- by tractable convex optimization. We further know that these methods are robust vis a vis additive noise and even outliers. In a different direction, researchers have developed computationally tractable methods for clustering high-dimensional data points that are assumed to be drawn from multiple low-dimensional linear subspaces. This talk will survey some exciting results in these areas.

Human infants learn spontaneously and effortlessly the language(s) spoken in their environments, despite the extraordinary complexity of the task. In the past 30 years, tremendous progress has been made regarding the empirical investigation of the linguistic achievements of infants during their first two years of life. In that short period of their life, infants learn in an essentially unsupervised fashion the basic building blocks of the phonetics, phonology, lexical and syntactic organization of their native language (see Jusczyk, 1987). Yet, little is known about the mechanisms responsible for such acquisitions. Do infants rely on general statistical inference principles? Do they rely on specialized algorithms devoted to language?
Here, I will present an overview of the early phases of language acquisition and focus on one area where a modeling approach is currently being conducted, using tools of signal processing and automatic speech recognition: the unsupervized acquisition of phonetic categories. It is known that during the first year of life, before they are able to talk, infants construct a detailed representation of the phonemes of their native language and loose the ability to distinguish nonnative phonemic contrasts (Werker & Tees, 1984). It will be shown that the only mechanism that has been proposed so far, that is, unsupervised statistical clustering (Maye, Werker and Gerken, 2002), does not converge on the inventory of phonemes, but rather on contextual allophonic units or subunits (Varadarajan, 2008). An information-theoretic algorithm wil be presented: it groups together allophonic variants based on three sources of information: the statistical distribution of their contexts, the phonetic plausibility of the grouping, and the existence of lexical minimal pairs (Peperkamp et al., 2006; Martin et al, submitted). It is shown that each of the three sources of information can be acquired  without presupposing the others. This algorithm is then tested on several natural speech corpora. 
The more general proposal is that early language bootrapping does not rely on learning principles necessarily specific to language. What is presumably unique to language though, is the way in which these principles are combined in a particular ways to optimize the emergence of linguistic categories after only a few months of unsupervized exposure to speech signals.  
 
Jusczyk, P. (1997). The discovery of spoken language. Cambridge, MA: MIT Press. 
Martin, A., Peperkamp, S., & Dupoux, E. (submitted).  Learning phonemes with a pseudo-lexicon.Maye, J., Werker, J., & Gerken, L. (2002). Infant sensitivity to distributional information can affect phonetic discrimination. Cognition, 82, B101–B111. Peperkamp, S., Le Calvez, R., Nadal, J.P. and Dupoux, E. (2006). The acquisition of allophonic rules: statistical learning with linguistic constraints. Cognition, 101, B31-B41Varadarajan, B., Khudanpur, S. & Dupoux, E. (2008). Unsupervised Learning of Acoustic Subword Units, in Proceedings of ACL-08: HLT, 165-168.  Werker, J.F., & Tees, R.C. (1984). Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant Behavior and Development, 7, 49-63.

This lab is designed for people who started recently to work in machine learning problems. In the first part of this lab, we will implement from scratch basic classification techniques and apply them to several datasets. In the second part, we will briefly review linear regression and then study and apply LWPR, a non-linear function approximation technique based on multiple local linear regressors.

This lab will present an hands-on (using matlab) on markov decision processes (markov processes, mdps, model estimation) and standard methods to solve them (VI, Q-Learing, ...). A second, more adavanced part will address inverse reinforcement learning and apprentiship learning.

Decision making is an important skill of autonomous agents. This lab course focuses on decision making under uncertainty in sensing and acting, common in many real-world systems. The associated planning problems can be formalized as partially observable Markov decision processes (POMDPs). This lab course will provide an introduction to the POMDP model and its solution methods, and the practical part will consists of implementing and testing different MDP and POMDP solving techniques. Finally, attention will be given to the problem of
decision making under uncertainty with multiple, interacting agents.

During the first part of the lab, we will overview some of the capabilities and methods to do learning and inference using Gaussian processes (GPs). A second, more advanced, part of the lab will cover the problems of active learning, experimental design, Bayesian optimization and submodular optimization based on GPs. During both labs, we will illustrate each part with realistic learning problems such as: robot kinematics, handwritten digits recognition or optimal sensor placement. Required software: Matlab or Octave.