 Priberam Machine Learning Lunch Seminar Speaker: Ramon Astudillo (INESC-ID) Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação) Date: Tuesday, March 6th, 2012 Time: 13:00 Lunch will be provided Title: Integration of Fourier Domain Speech Enhancement and Automatic Speech Recognition through Uncertainty Propagation Abstract: Speech enhancement techniques aim to recover the original clean signal underlying corrupted speech. Such techniques typically operate in the short-time Fourier transform (STFT) domain where phenomena like additivity of background noises, interfering speakers and echoes are easier to model. By contrast, automatic speech recognition (ASR), and in general most speech-related machine learning applications, operate on feature spaces that are non-linear transformations of the STFT. The reason for this is that such spaces provide a more compact representation of the acoustic space, the space of all acoustic realizations for a given task, and thus lead to simpler models. This talk discusses the integration of STFT speech enhancement and ASR using the concept of uncertainty propagation and decoding. This will include conventional speech enhancement in STFT domain, its associated uncertainty and various closed-form solutions for propagation into domains suitable for ASR.
 Priberam Machine Learning Lunch Seminar Speaker: Miguel Almeida (IST/UTL and Aalto University, Finland) Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação) Date: Tuesday, January 24th, 2012 Time: 13:00 Lunch will be provided Title: SSS: Separation of Synchronous Sources Abstract: The problem of separating synchronous sources (SSS) is a case of blind source separation (BSS) where independence of the sources is not satisfied. In SSS, the sources are assumed to be complex-valued, and different sources are phase-locked, which means that the relative phase lag between two sources is not uniform in [0,2*pi[. For this reason, the typical independent component analysis (ICA) tools are theoretically not applicable, and experiments show that they perform poorly in this task. In the SSS model, we assume that the phase lag between any two sources is constant. The only important assumption regarding the amplitudes of the sources is linear independence, although some nice results can be proven if the amplitudes are statistically independent. In this talk, I'll start by briefly discussing ICA, since it is relatively familiar in the Machine Learning community. I will then formulate the problem of SSS and detail the similarities and differences to the ICA problem. Afterwards, I will present two algorithms that were developed to tackle this problem, along with some nice theoretical properties of those algorithms. We will visit some very simple optimization problems and a little bit of complex algebra. Nothing complicated, I promise! I will finalize by presenting some simulated results, on 1) data which exactly follows the SSS model, and 2) data which deviates from the SSS model. -- Bio: Miguel Almeida is currently a joint PhD student at IST-UTL, Portugal, and at Aalto University (AU), Finland (formerly Helsinki University of Technology), under joint supervision of Prof. José Bioucas-Dias (IST), Prof. Ricardo Vigário (AU), and Prof. Erkki Oja (AU). He started his doctoral project in 2008 and spent the first two years of his PhD at AU. He has been at IST since 2010, and plans to finish his degree in the first semester of 2012. His PhD topic revolves around the SSS problem, and fits under the general topic of Machine Learning. More specifically, this project involves considerable amounts of Signal Processing and Optimization. Miguel holds an MSc in Physics and Technology Engineering (IST, 2006) and an advanced post-graduate degree in Biophysics (FC-UL, 2007).
 Priberam Machine Learning Lunch Seminar Speaker: Kalyanmoy Deb (http://www.iitk.ac.in/kangal/deb.htm) Host: Sara Silva (KDBIO/INESC-ID) Venue: IST Alameda, Sala Qa1.3 (Torre Sul) ***NOTE SPECIAL VENUE!*** Date: Monday, April 4th, 2011 ***NOTE SPECIAL DATE!*** Time: 12:00 ***NOTE SPECIAL TIME!*** Lunch will be provided Title: Innovization: Revealing Innovative Design Principles through Multi-Objective Optimization Abstract: Designing a component, process or a control system to achieve minimum or maximum of a single objective (or goal) often results in a single optimum solution describing the shape, dimensions, process or strategy of solving the task. Although such an optimized solution may already provide a new and innovative way of achieving the best objective value, it is almost never the case that practitioners are solely interested in a single objective. Moreover, a single optimum solution does not often provide adequate information to *learn* much about the problem, a matter which is ideally desired in engineering and scientific problem-solving tasks. In this seminar, we shall discuss an "innovization" procedure involving a multi-objective optimization algorithm for finding a set of trade-off optimal solutions. An investigation of such solutions is then expected to reveal useful design principles common to high-performing solutions. A number of interesting engineering case studies will be discussed to demonstrate how useful and innovative design principles can be deciphered by considering two or three-objective optimization problems. Such design innovations are difficult to achieve by any other means and the proposed systematic procedure should find a wide-spread applicability in the coming years. -- Bio: Kalyanmoy Deb is currently a Professor of Mechanical Engineering at Indian Institute of Technology Kanpur, India and is the director of Kanpur Genetic Algorithms Laboratory (KanGAL). He is the recipient of the prestigious Shanti Swarup Bhatnagar Prize in Engineering Sciences for the year 2005. He is a fellow of Indian National Science Academy (INSA), Indian National Academy of Engineering (INAE), Indian National Academy of Sciences (IASc), and International Society of Genetic and Evolutionary Computation (ISGEC). He has received Fredrick Wilhelm Bessel Research award from Alexander von Humboldt Foundation, Germany in 2003. His main research interests are in the area of optimization, optimal modeling and design and evolutionary algorithms. He has written two text books on optimization and more than 275 international journal and conference research papers. He is associated with 17 international journals. He has pioneered and a leader in the field of evolutionary multi-objective optimization. More information about his research can be found from http://www.iitk.ac.in/kangal/deb.htm.
 Priberam Machine Learning Lunch Seminar Speaker: André Lourenço (IT) Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação) Date: Tuesday, March 29th, 2011 Time: 13:00 Lunch will be provided Title: Towards a Finger Based ECG Biometric System Abstract: The electrocardiographic (ECG) signal has been shown to contain relevant information for human identification. Even though results validate the potential of these signals, data acquisition methods and apparatus explored so far compromise user acceptability, requiring the acquisition of ECG at the chest. In this talk we review the state of the art on using ECG as biometric trait. We present a finger based ECG biometric system, that uses signals collected at the fingers, through a minimally intrusive 1-lead ECG setup recurring to Ag/AgCl electrodes without gel as interface with the skin. The collected signal is significantly more noisy than the ECG acquired at the chest, motivating the application of feature extraction and signal processing techniques to the problem. -- Bio: André Lourenço is a phd student of Electrical and Computer Engineering at IST-IT (Instituto Superior Técnico – Instituto de Telecomunicações), under the supervision of prof. Ana Fred. He is also assistant professor at ISEL (Instituto Superior de Engenharia de Lisboa). His main research interests are pattern recognition and machine learning.
 Priberam Machine Learning Lunch Seminar Speaker: Jorge Marques (ISR) Venue: IST Alameda, Sala EA3 (Torre Norte) Date: Tuesday, March 15th, 2011 Time: 13:00 Lunch will be provided Title: Project ARGUS: Characterizing People Activities Using Multiple Motion Fields Abstract: Surveillance systems aim to characterize human activities and to detect abnormal behaviors. This task is specially challenging if the camera field of view is wide and the objects are far from the camera. In such operating conditions, it is not possible to extract detailed descriptions of the objects such as shape and color. In this case, most of the information is conveyed by the object trajectory and motion parameters. We therefore need to characterize trajectories and to be able to discriminate normal from abnormal behaviors. This talk presents a new representation for human activity analysis based on multiple motion fields, equipped with space-varying switching mechanisms. We will show that this description is flexible and intuitive. The model parameters have a meaning and they can be used to understand how people behave in a scene. Parameter estimation will be addressed using the EM method and several extensions will be discussed. -- Bio: Jorge Marques received the Ph.D degree and Aggregation title in Electrical and Computer Engineering from IST. He is currently with the department of Electrical and Computer Engineering, IST, where he is Associate Professor. His research interests are in the area of Image Processing and Pattern Recognition.
Priberam Machine Learning Lunch Seminar
Speaker: Andras Hartmann (INESC-ID)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Tuesday, November 2nd, 2010
Time: 13:00
Lunch will be provided
Title: "Mind The Gap: Reconstruction of missing cardiovascular signals using adaptive filtering"
Abstract:
In this talk I will introduce a robust method for filling in short missing segments in multiparameter Intensive Care Unit cardiovascular data. This work was inspired by the ``PhysioNet/Computing in Cardiology Challenge 2010: Mind the Gap''.
The interconnections between the signals were identified in the form of composite IIR transfer functions using the signals' history. A genetic algorithm was applied for inferring the filter coefficients. Assuming that the connections do not vary in time, we managed to reconstruct the missing signals using the yet available parallel measured signals and the transfer functions.
The results are promising, as this method achieved the 5th place among 53 participants of the challenge. We concluded that this approach can be efficient in reconstructing and even detecting missing or corrupted cardiovascular signals or other type of datasets with several modalities and strong interconnections between them.
--
Bio: András Hartmann received his MSc degree from Budapest University of Technology and Economics in Information Systems and Computational Engineering with specialization in software design in 2005. In 2008 he gained his MSc, this time in Biomedical Engineering in a joint program of Semmelweis Medical University and Budapest University of Technology and Economics. Since July 2009 he is a member of the INESC-ID KDBIO Group, working on the project DynaMo - Dynamical modeling, control and optimization of metabolic networks.
He is interested in modeling complex biological and physiological systems, in particular: identification and dynamic modeling of metabolic networks; spatial and temporal connectivity in human brain; and dynamic modeling of cardiovascular system.
Priberam Machine Learning Lunch Seminar
Speaker: Ricardo Vigário (Aalto University School of Science and Technology, Finland)
Venue: IST Alameda, Sala EA4 (Torre Norte)
Date: Friday, July 2nd, 2010
Time: 13:00
Lunch will be provided
Title: "From elements to networks of neuronal activity – a machine learning approach"
Abstract:
Neuroinformatics “combines neuroscience and the information sciences to develop and apply advanced tools for a major advancement in understanding the structure and function of the brain.” After introducing the speaker’s neuroinformatics research group, we will address issues related to the use and misuse of independent component analysis.
Departing from the traditionally simple evoked response paradigm, into the more natural neurocinematics one, also the neuronal responses are expected to take on rather complex network configurations. We will review two approaches to identify such communication strategies. In a functional magnetic resonance imaging setup, the first one is a hierarchical method, and assumes the existence of basic focal activation areas, which are combined to account for the complex neuronal responses.
Additional information is gathered directly from the stimuli. The second uses phase synchrony as the acting principle for the extraction of communication/control in high temporal resolution data, such as electro- and magnetoencephalograms.
Bio: Ricardo Vigário, D.Sc., is a docent and senior researcher at the Aalto University School of Science and Technology, Finland, where he teaches and leads a research group in Neuroinformatics. He has a basic degree in Applied Physics and a M.Sc. in Biomedical Engineering from the Faculty of Sciences of the University of Lisbon and a D.Sc. (tech) in Computer Science from the Helsinki University of Technology (current Aalto University), from 1992, 1994 and 1999, respectively. He has held a Marie Curie post-doctoral position in GMD – FIRST, Germany; was a visiting lecturer in Graz, Austria and Zaragoza, Spain; and a visiting associate professor in Grenoble, France. He was a pioneer in the independent component analysis of electrophysiological data. His fields of interest include statistical machine learning; the analysis of neuronal responses to natural stimuli; and various communication strategies within the central
nervous system.
 Priberam Machine Learning Lunch Seminar Speaker: João Graça (L2F, INESC-ID) Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação) Date: Tuesday, June 22th, 2010 Time: 13:00 Lunch will be provided Title: "Posterior Regularization Framework: Learning Tractable Models with Intractable Constraints" Abstract: Unsupervised Learning of probabilistic structured models presents a fundamental trade- off between richness of captured constraints and correlations versus efficiency and tractability of inference. In this thesis, we propose a new learning framework called Posterior Regulariza- tion that incorporates side-information into unsupervised estimation in the form of constraints on the model’s posteriors. The underlying model remains unchanged, but the learning method changes. During learning, our method is similar to the EM algorithm, but we solve a problem similar to Maximum Entropy inside the E-Step to enforce the constraints. We apply the PR framework to two different large scale tasks: Statistical Word Alignments and Unsupervised Part of Speech Induction. In the former, we incorporate two constraints: bijectivity and symme- try. Training using these constraints produces a significant boost in performance as measured by both precision and recall against manually annotated alignments for six language pairs. In the latter we enforce sparsity on the word tag distribution which is overestimated using the default training method. Experiments on six languages achieve dramatic improvements over state-of-the-art results. Bio: I am currently a 4th year PhD student (with MSc degree) in Computer Science Engineering at Instituto Superior Técnico, Technical University of Lisbon and a visiting student at University of Pennsylvania. My advisors are Luisa Coheur, Fernando Pereira and Ben Taskar. My main research interests are Machine Learning and Natural Language Processing. My current focus in on unsupervised learning with high level supervision in the form of constraints. I am a proud member of the Spoken Language Systems Lab (L2F) in Lisbon and of the Penn Research in Machine Learning (PRiML).
Priberam Machine Learning Lunch Seminar
Speaker: Matthijs Spaan (ISR)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Tuesday, June 8th, 2010
Time: 13:00
Lunch will be provided
Title: "Decision-theoretic Planning under Uncertainty for Active Cooperative Perception"
Abstract:
As robots leave research labs to operate more often in human-inhabited, larger environments, cooperation between sensor networks and mobile robots becomes crucial. For example, in urban scenarios, employing mobile robots is a need to augment the limited sensor coverage and improve detection and tracking accuracy. The fusion of sensory information between fixed surveillance cameras and each robot, with the goal of maximizing the amount and quality of perceptual information available to the system can be called cooperative perception. A promising decision-theoretic planning framework for active cooperative perception is that of Partially Observable Markov Decision Processes (POMDPs). The suitability of POMDPs for the previously depicted scenario arises from their ability to inherently trade off task completion, which could be react to a potential event that has been detected, and information gathering in a efficient way, that is decide to send a robot to improve situational awareness. In this talk we will discuss how planning under uncertainty can be applied to active cooperative perception problems.
--
Bio: Matthijs Spaan received a M.Sc. (2002) in AI and a Ph.D. (2006) in CS from the Universiteit van Amsterdam, The Netherlands. Currently he is a research scientist at the Institute for Systems and Robotics, Instituto Superior Técnico, Lisbon, Portugal, and he is the principal investigator of a national project on "Decentralized Planning Under Uncertainty for Cooperative Systems". His thesis was on "Approximate planning under uncertainty in partially observable environments" and was selected as a runner-up for EURON's 7th Georges Giralt PhD Award. His scientific interests are in planning under uncertainty, sequential decision making, autonomous robots, cooperative multiagent/multi-robot systems, (decentralized) partially observable Markov decision processes (POMDPs/Dec-POMDPs), reinforcement learning, machine learning and AI in general.
Speaker: Francisco Melo (INESC-ID)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Tuesday, May 25th, 2010
Time: 13:00
Lunch will be provided
Title: "GRADIENT APPROACHES TO REINFORCEMENT LEARNING"
Abstract:
In this talk I will present an overview of some of the past and current lines of research in reinforcement learning (RL), as well as some of the challenges that research in this area has faced in the last decades. I will describe a range of recent results that may bring significant advances on some of these fundamental research challenges, and yet rely on the "simplest" optimization approach - gradient search. The ultimate goal of this talk is to provide a high-level perception of RL while hint on current active avenues of research in this area.
--
Bio: Francisco S. Melo received his PhD in Electrical and Computer Engineering at Instituto Superior Técnico, in Lisbon, Portugal. During 2007 he held an appointment as a short-term researcher in the Computer Vision Lab, at the Institute for Systems and Robotics (Lisbon, Portugal) and in 2008 he joined the Computer Science Department of Carnegie Mellon University as a Post-Doctoral Fellow. Since June 2009 he is a Researcher at the Intelligent Agents and Synthetic Characters Group of INESC-ID, where he develops research within reinforcement learning, planning under uncertainty, multiagent and multi-robot systems, developmental robotics, and sensor networks.
Speaker: Xavier Anguera Miro
Venue: IST Alameda, Sala EA4 (Torre Norte)
Date: Friday, May 14th, 2010
Time: 13:00
Lunch will be provided
Title: "Multimodal pattern matching algorithms and applications"
Abstract:
After introducing myself and where I come from, in this talk I will focus on 3 projects I have been working in the last year. The first one is a novel pattern matching algorithm, based on the well known Dynamic Time Warping. The presented algorithm can be used to find real-valued subsequences within a longer sequence, without prior knowledge of their
start-end points. I have applied the algorithm for the task of acoustic matching, for which I will show some preliminary results. Then I will continue to explain a second DTW-based algorithm, this one being able do an online of two musical pieces. One of the music pieces can be input life or be retrieved from an audio file, while the second one is extracted from an online music video. The online alignment allows for the music video to be played in total synchrony with the corresponding ambient/recorded audio. Finally, I will talk about video copy detection, which is the task of finding video duplicate segments within a big database. I will explain our multimodal approach, based on audio-visual change-based features.
--
Bio: Xavier Anguera Miro: Ing. [MS] 2001 by UPC (Barcelona, Spain), [MS] 2001 European Masters in Language and Speech, Dr. [PhD] 2006 UPC University, with a thesis on speaker diarization for multi-microphone meeting recordings. From 2001 to 2003 he worked for Panasonic Speech Technology Lab in Santa Barbara, CA. From 2004 to 2006 he was a visiting researcher at the International Computer Science Institute (ICSI) in Berkeley, CA. Since 2007 he is with Telefónica Research in Barcelona, Spain working as a research scientist in the multimedi research group led by Dr. Nuria Oliver. Although his background is in acoustic analysis, in the last 3 years he has been very interested in the area of multimodal algorithms and applications.
Amanhã, mais uma edição.
Priberam Machine Learning Lunch Seminar
Speaker: Ruben Martinez-Cantin (http://users.isr.ist.utl.pt/~rmcantin/)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Thursday, April 29th, 2010
Time: 13:00
Lunch will be provided
Title: "Towards Closing the Loop: Active Learning for Robotics"
Abstract:
The ability to adapt to changing environment autonomously will be essential for future robots. While this need is well-recognized, most machine learning research focuses largely on perception and static data sets. Instead, future robots need to interact with the environment to generate the data that is needed to foster real-time adaptation based on all information collected in previous interactions and observations. In other words, we need to close the loop between the robot acting, robot sensing and robot learning. Novel active methods need to outperform passive methods by a margin that compensates the potential of the extra computational burden and the cost of the active data sampling.
In this talk, we present a common framework for active learning in different applications, such as planning, robot localization and mapping, calibration, sensor networks and computer graphics. Our results show that in many applications, active sampling provides an improvement, while in other applications is mandatory to achieve a good performance.
--
Bio: Ruben Martinez-Cantin is a postdoctoral researcher at the Insitute of Systems and Robotics at IST in Lisbon. Before joining IST, he received a PhD and MSc in Computer Science and Electrical Engineering from the University of Zaragoza in 2008 and 2003, respectively. During his PhD, he worked in the Robotics, Perception and Real Time Group under the supervision of Prof. José A. Castellanos in mobile robotics and Bayesian inference and reasoning. In 2006 and 2007, he has been a visiting scholar at the Laboratory of Computational Intelligence (LCI) at UBC in collaboration with Prof. Nando de Freitas. Previously, he worked as research assistant at University of Zaragoza, in vision based control for mobile robots and intelligent surveillance systems. He also developed some ideas for space robotics and got two grants by the European Space Agency.
His research interests include Bayesian inference, machine learning, robotics, computer vision and cognitive models. He is also interested in the popularization of science.
|
|
|
