Unconventional computing for Bayesian inference

Abstract: Since more than a century, perception in human and animals has been recognized to be an inference process. More recently, a number of experimental studies have shown that visual, vestibular and more generally multisensory perceptions could be very well accounted for by Bayesian models. However, it remains largely unclear how probability distributions are coded and how inferences are computed in the brain. We will discuss some of the proposed theories. In addition, we will extend these ideas by considering that simpler organisms, without complex nervous system, and even unicellular organisms exhibit also very well adapted sensory-motor behaviours in spite of a highly uncertain environment. We will then propose that Bayesian computing might be a very fundamental capacity of living organisms and we will show how this capacity could be implemented by biochemical networks.

ISIR, UPMC, Paris, coordinator of EU FET project BAMBI – Bottom-up Approaches to Machines dedicated to Bayesian Inference.

short bio: Jacques DROULEZ (born in 1950) received a mathematical and engineer training (Ecole Polytechnique, Paris), a complete medical training (MD: Lariboisière - St Louis Hospital, Paris), a master in biochemistry and a Habilitation to supervise research in cognitive sciences. He has got a fellowship from the Centre National d’Etudes Spatiales (1978-1982). He is now Director of Research at the Centre National de la Recherche Scientifique (CNRS), and head of the research team « active perception and probabilistic approach » previously at the Laboratory of Physiology of Perception and Action (CNRS - Collège de France) and now at ISIR-UMPC. His main research themes are the perception of 3D motion and objects, the theoretical study of models for multi-sensory interactions and the adaptive motor control. He has about 100 publications in international journals including one in PNAS on sensory-motor integration model and one in Nature on object perception during self-motion. He is involved in several European and national research programs and in multidisciplinary scientific networks.

Abstract: In this presentation we will explain how to use Bayesian Programming to develop applications in robotics. After a short introduction to the main features of Bayesian Programming, we will present examples to illustrate more specifically how to learn reactive behaviours, how to implement filters, and how to build hierarchical models. We will introduce in more details the original concept of coherence variables to deal with multi inference paths and to reason with soft evidences. Finally, we will discuss some hints on possible development of specific probabilistic hardware for robotics.

ISIR, UPMC, Paris, author of “Bayesian Programming” and “Probabilistic Reasoning and Decision Making in Sensory-Motor Systems” .

short bio:Pierre BESSIERE (born in 1958) is a senior researcher at CNRS (Centre National de la Recherche Scientifique) since 1992. He took his engineering degrees (1981) and his PhD (1983) in computer science from INPG (Institut National Polytechnique de Grenoble). He did a Post-Doctorate at SRI International (Stanford Research Institute) working on a project for NASA (National Aeronautics and Space Agency). He then worked for five years in an industrial company as the leader of different artificial intelligence projects. Since he came back to research in 1989, his main research concerns have been evolutionary algorithms and probabilistic reasoning for perception, inference and action. He led the Bayesian Programming research group (Bayesian-Programming.org) on these two subjects in Grenoble before moving to Paris at LPPA. Fifteen PhD diplomas and numerous international publications are fruits of the activity of this group during the last 15 years. He also led the BIBA (Bayesian Inspired Brain and Artefact) and was a Partner in the BACS (Bayesian Approach to Cognitive Systems) European project. He is a co-founder and scientific adviser of the ProbaYes Company, which develops and sells Bayesian solutions for the industry.

Abstract: How should the uncertainty and incompleteness inherent to sensing the environment be represented and modelled in a way that will increase the autonomy of a robot? How should a robotic system perceive, infer, decide and act efficiently? These are two of the challenging questions the robotics community and robotic researchers have been facing. In this presentation we will address how probabilistic techniques and Bayesian inference can be used to effectively tackle these issues in the design of more autonomous, more intelligent and adaptive artificial systems.

ISR - University of Coimbra, author of "Probabilistic Approaches for Robotic Perception" .

short bio:João Filipe Ferreira (born in 1973) has been an Invited Assistant Professor at the University of Coimbra since 2011. He received the B.Sc. (five-year course), M.Sc., and Ph.D. degrees in electrical engineering and computers from the University of Coimbra, Coimbra, in 2000, 2005, and 2011, respectively. His current main research interests are spread out through three broad scientific themes: Artificial Cognition, Probabilistic Modelling and Autonomous Systems. Within these themes, the following topics receive his main focus: bioinspired perception, navigation and cognition, autonomous robotics, social robotics and human-robot interaction. However, his research interests are not limited to these subjects: over the years, he has also produced contributions in medical image processing and 3D scanning. He is the main author of the 2014 textbook "Probabilistic Approaches for Robotic Perception" (Springer STAR series). He has been a staff researcher at the ISR since 1999 (integrated member since 2011), and a member of the IEEE and the IEEE Robotics and Automation Society (RAS) since 2012 (Treasurer in the Portuguese Chapter since 2014), the IEEE Life Sciences Community since 2013, the IEEE Systems, Man, and Cybernetics Society since 2015 and the IEEE Computational Intelligence Society since 2015. He has participated in several important European projects over the years, and is currently a staff researcher for the ISR team on the European Integrated Project ``BAMBI -- Bottom-up Approaches to Machines dedicated to Bayesian Inference'' (FET Project - FP7-ICT-2013-C), from January 2014 to December 2016. He is also the Principal Investigator (PI) for the FCT/COMPETE nationally funded project CASIR (Coordinated Control of Stimulus-Driven and Goal-Directed Multisensory Attention Within the Context of Social Interaction with Robots - PTDC/EEI-AUT/3010/2012), running from April 2013 to July 2015.

Abstract: How should the uncertainty and incompleteness inherent to sensing the environment be represented and modelled in a way that will increase the autonomy of a robot? How should a robotic system perceive, infer, decide and act efficiently? These are two of the challenging questions the robotics community and robotic researchers have been facing. In this presentation we will address how probabilistic techniques and Bayesian inference can be used to effectively tackle these issues in the design of more autonomous, more intelligent and adaptive artificial systems.

ISR - University of Coimbra, author of "Probabilistic Approaches for Robotic Perception”" .

short bio:Jorge Dias is professor at the University of Coimbra (UC) and Khalifa University, Abu Dhabi (KU) and researcher at ISR. His current research areas are computer vision and robotics, with activities and contributions in these fields since 1984. He is currently PI in the Social Robot project (FP7 (Marie Curie), no.:285870), and has been the main researcher for the ISR team in several projects financed by the European Commission (FP 6 and 7) BACS - Bayesian Approach to Cognitive Systems (FP6-IST-027140 Feb2006-Feb2010); IRPS - Intelligent Robotic Porter System (EU-IRPS FP6-IST-45048 an2007-Jan2011); PROMETHEUS - Prediction and interpretation of human behaviour based on probabilistic structures and heterogeneous sensors (FP7 – 214901 Aug2008-Aug2011); HANDLE - Developmental pathway towards autonomy and dexterity in robot in-hand manipulation (FP7-2008– 231640 Feb2009- Feb2013), across which there was a focus on applying Bayesian techniques to artificial perception in robotics. He is currently local PI for the ISR team on the European Integrated Project ``BAMBI -- Bottom-up Approaches to Machines dedicated to Bayesian Inference'' (FET Project - FP7-ICT-2013-C).

Abstract: Markov Chain Monte Carlo (MCMC) based methods have been the main tool used by practitioners and researchers for Bayesian Inference for over 20 years due to their flexibility and theoretical properties that guarantee unbiased results. Nevertheless, these methods pose computational challenges when target complex problems. In this talk, we will expand on state-of-the-art MCMC methodology with a view on the computational challenges and the hardware platform trends of the next decade, aiming to design MCMC computational platforms capable to address large-scale inference problems utilising complex models. More specifically, we will discuss how reconfigurable computing can be utilised for such problems, as well as we will see some of the current challenges towards this direction.

Imperial College London, United Kingdom, general chair and pc of various conferences in the field of reconfigurable computing, editorial board member of IET Computers and Digital Techniques and of Journal of Systems Architecture. "home page”" .

short bio: Dr. Christos Bouganis received the M.Eng degree in Computer Engineering and Informatics from University of Patras Greece in 1998, the MSc degree in Communications and Signal Processing in 1999 and the Ph.D. degree in 2004 both from Imperial College London. He joined the Department of Electrical and Electronic Engineering as academic faculty in 2007. He is currently a Senior Lecturer within Department of Electrical and Electronic Engineering and is also the Director of the MSc in Analogue and Digital Integrated Circuit Design. His research inlcudes the theory and practice of reconfigurable computing and design automation, mainly targeting digital signal processing algorithms. His work is currenlty focused on Computer Vision and Image Processing, Machine Learning, Markov Chain Monte Carlo Systems, and computing with unreliable harwdare. He currently serves on the program committees of many international conferences, including FCCM, FPL, FPT, DATE, SPPRA, and VLSI-SoC and is an editorial board member of IET Computers and Digital Techniques and Journal of Systems Architecture.