Fuzzy Systems in Health Care
Susana M. Vieira, University of Lisbon, Portugal
Differentiability in Quadratic Mean of Deep Networks
Joseph Rynkiewicz, Université de Paris 1 Panthéon-Sorbonne, France
Recent Research Topics in Evolutionary Multiobjective Optimization: A Personal Perspective
Carlos Coello Coello, CINVESTAV-IPN, Mexico
Machine Learning in Non-Stationary Environments
Barbara Hammer, Bielefeld University, Germany
Fuzzy Systems in Health Care
Susana M. Vieira
University of Lisbon
Portugal
Brief Bio
Susana Vieira received the MSc and PhD degrees both in Mechanical Engineering (Mec. Eng.) in 2005 and 2010, respectively, from Instituto Superior Técnico (IST), University of Lisbon, Portugal. She was a Teaching Assistant at IST, in Mec. Eng., from 2005 to 2006, in 2009, she was an Invited Teacher at the Erasmus University of Rotterdam, the Netherlands, and in 2014 she was a visiting Scholar at MIT, Boston, USA. From 2012 to 2018 she was an invited assistant professor at IST, University of Lisbon and since November 2020 she is an associate professor at IST, University of Lisbon and a Senior researcher at the Center of Intelligent Systems (CIS), IDMEC- IST since 2012. Her main research area is Soft Computing and Artificial Intelligence, more specifically she works in machine learning, feature selection, fuzzy modelling, fuzzy optimization and metaheuristics. Her research focuses mainly on the development of computational intelligence methods for knowledge data discovery. Recently these methods are being used to identify important factors or features that lead to unfavourable or favourable clinical conditions of patients in Intensive Care Units, and design specific decision models that support clinicians’ decisions.
Abstract
Data accumulates in a speed unmatchable by human capacity of data processing. Thus, approximation of unknown functions from sampled data is an important activity in modern modeling and systems theory. It is important to develop models from data, which have sufficient generalization power and can describe the underlying process with accuracy, despite the nonlinearity and the complexity of these processes. Fuzzy systems are known to be universal function approximators that can be used to model real-world systems. Further, they have the additional advantage that the generated models can be interpreted in linguistic terms.
Health care is one of the areas where data has been growing exponentially. Improving quality, safety or clinical effectiveness as well as reducing costs are nowadays the main concerns for health care decision-makers. These are challenging problems and the structure or design of a system may influence the outcome, and is likely to be more significant in high-acuity, complex environments such as intensive care units (ICUs). Patients here are among the sickest patients in the hospital, and decisions that are made can literally mean the difference between life and death.
In clinical decision support systems, it is crucial to interpret the developed models by: determining which attributes are chosen by the artificial intelligence techniques and what is their clinical significance; propose alternative procedures or develop criteria for classifying patients into patient sub-group; and designing a post-implementation assessment of how well the system meets the goals.
Fuzzy modelling can have an important role in health care as it can provide a transparent description of the system that reflects the nonlinearity of the system. The rule-based nature of the models allows for a linguistic description of the knowledge captured in the model. It can also help the identification of important factors or features that identify specific groups of patients within a specific clinical setting, The design of specific decision models, will support clinicians’ decisions in terms of identifying the most suitable therapy for a specific patient in the ICU, in order to achieve more favourable clinical outcomes and preventing poor outcomes due to practice variation.
Differentiability in Quadratic Mean of Deep Networks
Joseph Rynkiewicz
Université de Paris 1 Panthéon-Sorbonne
France
Brief Bio
Joseph Rynkiewicz was born on February 10, 1968, in Antony (France). He obtained his Ph.D. in applied mathematics from the Université Paris-I Panthéon Sorbonne in 2000. He is "Maître de Conférences" at the Université Paris-I and a member of the SAMM research center. His research on asymptotic properties of the statistical estimator and practical applications of new statistical tools like artificial neural networks, hidden Markov models, and Bayesian networks. His publications range from asymptotic studies of models under loss of identifiability to forecast ozone peaks and exceedance levels using neural classifiers and weather predictions. He also worked on industrial applications like the speed of wind for the wind farm project. He teaches theoretical statistics, data mining, survey, and Deep learning from license to Master at the Université Paris-I.
Abstract
Deep neural networks are parametric functions whose parameters are optimized to minimize a cost function, often the opposed of the log-likelihood (the cross-entropy). Even if the dimension of the parameter vector is huge and the distribution law of the estimated parameters is rarely used, it is still interesting to know the local properties of the model. Indeed, a very well initialize network will work much better than random initial parameters. This property is routinely used when a big model estimated on an extensive set of data is finetuned on a small data set. This technic is called "transfer learning".
We propose to explain why transfer learning works so well by showing that the log-likelihood of the network around a suitable parameter is similar to the log-likelihood of a normal location model and very smooth. Such property, known as local asymptotic normality, is a consequence of the differentiability in the quadratic mean of the model.
We will give assumptions for the differentiability in quadratic mean of Deep Network with ReLU activation functions even such model is not differentiable in numerous points. Finally, we will show that such assumptions are checked in classical frameworks where these models are used, like images classification and natural language processing.
Keywords: Local asymptotic normality, differentiability in quadratic mean, Deep networks.
Recent Research Topics in Evolutionary Multiobjective Optimization: A Personal Perspective
Carlos Coello Coello
CINVESTAV-IPN
Mexico
http://delta.cs.cinvestav.mx/~ccoello
![]()
Brief Bio
Carlos Artemio Coello Coello received a PhD in Computer Science from Tulane University (USA) in 1996. His research has mainly focused on the design of new multi-objective optimization algorithms based on bio-inspired metaheuristics (e.g., evolutionary algorithms), which is an area in which he has made pioneering contributions. He currently has more than 500 publications, including more than 160 journal papers and 50 book chapters. He has published a monographic book and has edited 3 more books with publishers such as World Scientific and Springer. He has supervised 22 PhD theses (including 3 in Argentina) and 48 Masters thesis (including one in France). Several of the PhD theses that he has supervised, have received awards in national competitions. He has also received (with his students) several “best paper awards” at different international conferences. He is also the only Latin American who has been awarded (twice) the “outstanding paper award” of the IEEE Transactions on Evolutionary Computation. His publications currently report 55,804 citations in Google Scholar. According to Scopus, Dr. Coello has 22,010 citations, excluding self-citations and citations from all his co-authors. His h-index is 95, according to Google Scholar, 65 according to Scopus and 60 according to the ISI Web of Science. In the ShanghaiRanking’s Global Ranking of Academic Subjects 2016 developed by Elsevier, he appears as one of the 300 most highly cited scientists in the world in “Computer Science”, occupying the first place in Mexico. He has received several awards, including the National Research Award (in 2007) from the Mexican Academy of Science (in the area of exact sciences), the 2009 Medal to the Scientific Merit from Mexico City's congress, the Ciudad Capital: Heberto Castillo 2011 Award for scientists under the age of 45, in Basic Science, the 2012 Scopus Award (Mexico's edition) for being the most highly cited scientist in engineering in the 5 years previous to the award and the 2012 National Medal of Science in Physics, Mathematics and Natural Sciences from Mexico's presidency (this is the most important award that a scientist can receive in Mexico). He also received the Luis Elizondo Award from the Tecnológico de Monterrey in 2019. Additionally, he is the recipient of the 2013 IEEE Kiyo Tomiyasu Award, "for pioneering contributions to single- and multiobjective optimization techniques using bioinspired metaheuristics", of the 2016 The World Academy of Sciences (TWAS) Award in “Engineering Sciences”, and of the 2021 IEEE Computational Intelligence Society Evolutionary Computation Pioneer Award. Since January 2011, he is an IEEE Fellow. He is currently the Editor-in-Chief of the IEEE Transactions on Evolutionary Computation. He is Full Professor with distinction (Investigador Cinvestav 3F) at the Computer Science Department of CINVESTAV-IPN in Mexico City, Mexico.
Abstract
In this talk, some research topics that are worth exploring (from the personal perspective of the speaker) in evolutionary multi-objective optimization will be briefly discussed. Such topics include scalability of multi-objective evolutionary algorithms (both in objective and in decision variable space), indicator-based selection, hyper-heuristics, parallelism and scalarizing functions. In the final part of the talk, some personal thoughts about the future of the field will also be briefly discussed.
Machine Learning in Non-Stationary Environments
Barbara Hammer
Bielefeld University
Germany
Brief Bio
Barbara Hammer received her Ph.D. in Computer Science in 1995 and her venia legendi in Computer Science in 2003, both from the University of Osnabrueck, Germany. From 2000-2004, she was leader of the junior research group 'Learning with Neural Methods on Structured Data' at University of Osnabrueck before accepting an offer as professor for Theoretical Computer Science at Clausthal University of Technology, Germany, in 2004. Since 2010, she is holding a professorship for Theoretical Computer Science for Cognitive Systems at the CITEC cluster of excellence at Bielefeld University, Germany.
Her areas of expertise include hybrid systems, self-organizing maps, clustering, and recurrent networks as well as applications in bioinformatics, industrial process monitoring, or cognitive science. She is leading the task force 'Data Visualization and Data Analysis' of the IEEE CIS Technical Committee on Data Mining, and the Fachgruppe Neural Networks of the GI.
Abstract
One of the main assumptions of classical machine learning is that data are generated by a stationary concept. This is often violated in practice if unexpected events happen, sensor drift is present, of models should be individualized for a specific user, for example. In such situations, a number of challenges occur: How to detect that drift occurs? How to adapt the model on the fly, given new data, without catastrophic forgetting? How to incorporate measures which can mediate suboptimal behavior? In the talk, we will address three algorithmic challenges: flexible drift detection with minimum requirements, life-long model adaptation, and reject options. We will demonstrate the behavior of the algorithmic models in the context of medical applications and industry4.0, respectively.