Confirmed plenary talks
James C. Bezdek
Title: Streaming Clustering is not Clustering!
This talk concerns models and algorithms that are generally described as “streaming clustering”. Some of the semantics and methods that are used in this field are co-opted from static (batch) clustering, but often, they don’t serve their purposes for streaming data very well. A review of “state of the art” methods such as sequential k-means, Birch, Clustream, Denstream, etc. shows that methods borrowed from classical batch techniques don’t transfer well to the streaming data case. Most of these models fail to acknowledge that the data are seen but once in real streaming analysis (e.g., intrusion detection). When the data are not saved, batch clustering ideas such as pre-clustering assessment, partitioning, and cluster validity are not relevant.Several current algorithms are briefly reviewed and illustrated (albeit poorly, with small labeled data sets!). Then I will discuss two new ideas: (i) incremental Stream Monitoring Functions that “see” some of the things your streaming algorithm is doing; and (ii) a new approach for visual assessment of streaming data, which displays the visual structure that the iVAT algorithm thinks it sees in the data stream. The conclusions? Useful analysis of real streaming data is in its infancy. We need to carefully define the objectives of streaming analysis, and then choose terminology and methods that suit this evolving paradigm.
Jim received the PhD in Applied Mathematics from Cornell University in 1973. Jim is co-founder with Enrique Ruspini of NAFIPS; is a past president of NAFIPS (North American Fuzzy Information Processing Society), IFSA (International Fuzzy Systems Association) and the IEEE CIS (Computational Intelligence Society): founding editor the Int’l. Jo. Approximate Reasoning and the IEEE Transactions on Fuzzy Systems: Life fellow of the IEEE and IFSA; and a recipient of the IEEE 3rd Millennium, IEEE CIS Fuzzy Systems Pioneer, IEEE Rosenblatt and Kampe de Feriet Awards. Jim’s interests: clustering in big data, woodworking, optimization, cigars, data visualization, fishing, anomaly detection, blues music, streaming data analysis, and poker. Jim retired in 2007, and will be coming to a university near you soon.
Title: Teaching a fuzzy logic based eXplainable AI class at the University of Cincinnati
In recent times, soft computing techniques are providing effective solutions to large scale, complex problems in a variety of fields including mobile and networked robotics, autonomous vehicles transportation & traffic management as well as predictive modeling for personalizing medical treatment. In Fall 2019, a new graduate level class was offered with the intention of providing students with a introduction to “eXplainable AI” and soft computing using fuzzy logic and the genetic fuzzy based hybrid intelligent systems. Students enrolled from several departments across the university including aerospace, civil, and mechanical engineering; computer science; psychology, and design, art, architecture and planning. Student learning outcomes include: (1) To appreciate the strengths of soft computing systems and its impact on the growing field of artificial intelligence. (2) To acquire a basic understanding for development of a fuzzy logic rule base, the fuzzy inference system and its interfaces (fuzzification and defuzzification) and perform fuzzy reasoning using them. (3)To possess a working knowledge of bio-inspired genetic algorithms. (4) To design a genetic fuzzy system for engineering application. (5) To creatively channel genetic fuzzy AI developments into the development of a meaningful final project. In this plenary talk, we survey the challenges faced in developing this class and the approach to achieve the student outcomes. Students were assessed over twelve deliverables spread throughout the semester. Finally, we reflect upon student feedback, the lessons learned and the modifications to the learning experience when the class is taught again in Fall 2020.
Prof. Kelly Cohen is the Brian H. Rowe Endowed Chair and Interim Head, Department of Aerospace Engineering and Engineering Mechanics at the University of Cincinnati in Ohio. He received his PhD. in Aerospace Engineering from the Technion, Israel Institute of Technology, in 1999 applying fuzzy logic control to aerospace applications. His main expertise lies in the area of intelligent systems and unmanned aerial systems (UAS). He has utilized genetic fuzzy logic-based algorithms to develop decision support systems for autonomous collaborating robotics as well as predictive modeling for personalizing medical treatment in neurological disorders. Dr. Cohen joined the University of Cincinnati in 2007 after working for four years at the USAF Academy in Colorado. Since 2010, Dr. Cohen has graduated 9 PhD students and 24 MS students and secured around 4.0 M$ in research funding from a variety of agencies such as NSF, NIH, USAF, DHS, ODOT and NASA. He has over 90 per reviewed archival publications, and another 350 conference papers/presentations/invited seminars.
Martine De Cock
Title: Privacy-preserving machine learning
The potential of machine learning to influence and to improve the quality of our daily lives continues to grow. Many machine learning applications rely on personal user data. There is a perceived tension between the desire to protect data privacy on one hand, and to promote an economy based on free-flowing data on the other hand. This tension can be alleviated by techniques for computation over encrypted data, allowing machine learning models to be trained and deployed without any party ever needing to disclose their data. We illustrate with various applications how to balance the tradeoff between security (no information leakage), accuracy, and efficiency in privacy-preserving machine learning based on Secure Multiparty Computation.
Martine De Cock is a Professor at the School of Engineering and Technology, University of Washington Tacoma. She has over 190 peer reviewed publications in international journals and conferences on artificial intelligence, data mining, machine learning, information retrieval, web intelligence, and logic programming. Her current research interests are privacy-preserving machine learning (PPML) and cybersecurity. Her team won Track IV of iDASH2019, the most significant competition for privacy-preserving analysis of genomic data in the world, with over 64 different teams from 17 countries competing in the track.
Title: Bias and Algorithmic Discrimination in Machine Learning
AI and machine learning tools are being used with increasing frequency for decision making in domains that affect peoples’ lives such as employment, education, policing and loan approval. These uses raise concerns about biases and algorithmic discrimination and have motivated the development of fairness-aware mechanisms in the machine learning (ML) community. In this talk, I will show how to measure bias and define fairness and why this is a challenging task. Then, I will present three categories of techniques to ensure fairness in different stages of the ML pipeline. I will conclude my talk with takeaways, open questions and future directions towards building a trustworthy ML system.
Golnoosh Farnadi obtained her PhD from KU Leuven and UGent in 2017. During her PhD, she addressed several problems in user modeling by applying and developing statistical machine learning algorithms. She later joined the LINQS group of Lise Getoor at UC Santa Cruz, to continue her work on learning and inference in relational domains. She is currently a post-doctoral IVADO fellow at UdeM and Mila, working with professor Simon Lacoste-Julien and professor Michel Gendreau on fairness-aware AI. She will be joining the decision sciences department at HEC Montréal as an assistant professor this Fall. She has been a visiting scholar at multiple institutes such as UCLA, University of Washington Tacoma, Tsinghua University, and Microsoft Research, Redmond. She has had successful collaborations that are reflected in her several publications in international conferences and journals. She has also received two paper awards for her work on statistical relational learning frameworks. She has been an invited speaker and a lecturer in multiple venues and the scientific director of IVADO/Mila “Bias and Discrimination in AI” online course.
Title: Rule Mixtures and Fuzzy Cognitive Maps: Advances in Feedforward and Feedback Fuzzy Systems
Probability is the language of modern machine learning. This makes it all too easy to overlook the expressiveness and utility of fuzzy sets and systems. A new result shows that additive fuzzy systems combine the expressive power of fuzzy sets with the mathematical machinery of probability theory because a generalized mixture probability density completely characterizes such additive systems. This mixture result does not apply to non-additive or min-max systems. Rules correspond to mixed likelihood functions. This gives a Bayesian posterior density over
all rule firings along with a conditional-variance measure of the output certainty?instant explainable AI (XAI). Mixing mixtures further allows arbitrary rulebase fusion. Feedback fuzzy cognitive maps also mix to form scalable causal networks. They trade the numerical precision of probabilistic DAGs or Bayesian belief trees for pattern prediction, faster and scalable computation, ease of combination, and richer feedback representation.
Dr. Bart Kosko is a professor of electrical and computer engineering, and law, at the University of Southern California. He holds degrees in philosophy, economics, applied mathematics, electrical engineering, and law. Dr. Kosko has published the textbooks Neural Networks and Fuzzy Systems, and Fuzzy Engineering, the trade books Fuzzy Thinking, Heaven in a Chip, and Noise, the edited volume Neural Networks for Signal Processing, the co-edited volume Intelligent Signal Processing, and the novels Nanotime and Cool Earth.
Title: Deep Learning (Partly) Demystified
Successes of deep learning are partly due to appropriate selection of activation function, pooling functions, etc. Most of these choices have been made based on empirical comparison and heuristic ideas. In this paper, we show that many of these choices — and the surprising success of deep learning in the first place — can be explained by reasonably simple and natural mathematics.
Vladik Kreinovich is Professor of Computer Science at the University of Texas at El Paso. His main interests are representation and processing of uncertainty, especially interval computations and intelligent control. He has published nine books, 30 edited books, and more than 1,600 papers. Vladik is Vice President of the International Fuzzy Systems Association (IFSA), Vice President of the European Society for Fuzzy Logic and Technology (EUSFLAT), Fellow of International Fuzzy Systems Association (IFSA), Fellow of Mexican Society for Artificial Intelligence (SMIA), Fellow of the Russian Association for Fuzzy Systems and Soft Computing. He is Treasurer of IEEE Systems, Man, and Cybernetics Society.
Title: Reproducibility with Microsoft Research Open Data
Access to repositories with open data sources is critical for reproducibility of research. Microsoft Research Open Data is a unique initiative that combines features of a traditional data repository with easy access to compute resources. The main aim is to increase reproducibilty of research outcomes by making datasets associated with research papers published by Microsoft researchers available broadly. Such datasets are hosted along with relevant metadata to make it easier to discover related assets that aid reproducibilty. The repository is hosted in the cloud and has seamless integration to Azure compute enabling users to run experiments on the data using convenient cloud compute resources, thus alleviating the signifiant costs and bandwidth constraints incurred in moving the data. The reproducibilty aspect within Microsoft Research Open data is framed using three pivots – Investment, Incentive and Infrastructure.
Vani Mandava is a Director of Data Science at Microsoft Research at Redmond with over a decade of experience in engineering teams designing and shipping software that is in use by millions of users across the world. She is passionate about enabling academic researchers and institutions develop technologies that fuel data-intensive scientific research using advanced techniques in data management, data mining, especially leveraging Microsoft cloud and AI platform. She has enabled the adoption of data mining best practices in various v1 products across Microsoft client, server and services in MS-Office, Sharepoint, Online Services (Bing Ads) organizations and in the Academic Search service. She co-authored a book ‘Developing Solutions with Infopath’ and holds patents in service engineering. She leads Microsoft Research Outreach efforts at Data Science Institutions in US universities. She co-chaired KDD Cup 2013 and partners with many academic and government agencies including NSF supported Big Data Innovation hub effort, a consortium coordinated by top US data scientists and expected to advance data-driven innovation nationally in the US. She currently leads microsoftopendata.com project, a cloud based open data repository.
Title: Mathematics in Deep Neural Networks
- It is well known that the first layers of DNN retrieve primitive features, such as local means, edges, etc. However, this knowledge is not applied in the assignment of weights. Still initialization with random Gaussian kernels dominates. We propose to brake down this tradition.
- We offer: initialization with predefined kernels whose weights are assigned and whose semantic meaning is known.
- We show stability – NN does not significantly change pre-assigned kernels.
- For theoretical justification, we use the theory of F-transforms of higher order.
- In the literature, only the approximation property of DNNs is implied. However, it is important to find an approximation (in the attribute space) that promotes separation.
- When formulating DNN as an algorithmic tool, the input data is simply n-dimensional points. This does not help in extracting the object, because the points have no structure. Therefore, it’s better to talk about functionalityof inputs.
Professor Irina Perfilieva, Ph.D., received the degrees of M.S. (1975) and Ph.D
(1980) in Applied Mathematics from the Lomonosov State University in Moscow,
Russia. At present, she is full professor in the University of Ostrava (CZ) and a Head of Theoretical Research Department in the Institute for Research and Applications of Fuzzy Modeling (IRAFM). She is: the author and co-author of six books on mathematical principles of fuzzy sets, fuzzy logic and their applications; area editor of IEEE TFS, International Journal of CIS (IJCIS), Knowledge-Based Systems, editorial board member of the journals: Fuzzy Sets and Systems, Journal of Uncertain Systems, Journal of Intelligent Technologies and Applied Statistics, Fuzzy Information and Engineering. She works as a member of Program Committees of the most prestigious International Conferences and Congresses in the area of fuzzy and knowledge-based systems. She received the memorial Da Ruan award at FLINS 2012 and the best paper award at IFSA 2019. She is an EUSFLAT Honorary Member, and the IFSA Fellow. She received a special price at the 2010 Seoul International Invention Fair and several awards for the best paper, IFSA 2019 as the last. She has two patents. Her scientific interests lie in the area of applied mathematics and mathematical modeling where she successfully uses modern as well as classical approaches. For the past five years, she has been working in the field of image processing, pattern recognition and the theoretical foundations of deep learning.