Tutorials

To access the open-access Sunday Program, click here: https://fuzz-virtual.org/schedule

To help attendees choose which tutorials may be most suitable for them, each tutorial below provides a ‘chilli-rating’ as follows:
🌶: Suitable for undergraduate and interested high school students and equivalent, no prior expertise required.
🌶🌶: Most useful to attendees with some background in X.
🌶🌶🌶: A well-developed understanding of X  is required to benefit fully from this tutorial.

A Hands-on Tutorial on Apache Spark for Solving Data Science Problems with Fuzzy logic

🌶: Some background on programming in Python is required, and knowing something about fuzzy modelling.
Organisers:  Isaac Triguero, Alberto Fernández, Mikel Galar.

A Top-Down Approach to Rule-Based Fuzzy Systems

🌶🌶: Most useful to attendees with some background in rule-based fuzzy systems
Organiser: Jerry Mendel.

Besides fuzzy sets: new theories for promptly moving your crisp systems’ full power to the fuzzy side.

🌶🌶: Background on Fuzzy Set Theory basics makes it easier to understand the tutorial.
Organiser: Daniel Sanchez.

Building interpretable fuzzy inference systems in Python

🌶🌶: A little prior knowledge about python and fuzzy logic is assumed; also, attendees are expected to already know why “explainable” systems are useful.
Organisers: Uzay Kaymak, Marco Nobile, Simone Spolaor, Caro Fuchs.

Explainable Fuzzy Systems: Paving the way from Interpretable Fuzzy Systems to Explainable Artificial Intelligence

🌶🌶: Most useful to attendees with some background in Explainable Artificial Intelligence.
Organisers: Jose Alonso, Ciro Castiello, Corrado Mencar, Luis Magdalena.

Fuzzy Systems for Brain Sciences & Interfaces.

🌶: No prior expertise required.
Organisers: Javier Andreu-Perez, Ching Teng Lin.

How big is too big? Clustering in (static) BIG DATA with the Fantastic 4

🌶🌶: TBC
Organiser: James Bezdek.

New Optimization Techniques for TSK Fuzzy Systems: Classification and Regression

🌶🌶: Most useful to attendees with some background in fuzzy system training.
Organiser: Dongrui Wu.

Recent Developments in Artificial Intelligence and Fuzzy Systems

🌶🌶: Most useful to attendees with some background in Artificial Intelligence and Fuzzy Systems.
Organisers: Alexander Gegov, Uzay Kaymak, Joao Sousa.

Towards Trusting AI/ML Decision-Making of Autonomous Systems <Cancelled>

🌶🌶: Most useful to attendees with some background and understanding of AI/ML and fuzzy set theory being helpful.
Organiser: Stanton Price.

Uncertainty modeling in adversarial learning

🌶🌶: Most useful to attendees with some background in machine learning and fuzzy sets.
Organiser: Xi-Zhao Wang.

Capture and Handling of Uncertainty at Source – Using Intervals in AI and why it matters

🌶: While familiarity with data analysis and statistics (regression specifically) is valuable, no specific prior knowledge is required to benefit from the tutorial.
Organisers: Christian Wagner, Vladik Kreinovich, Shaily Kabir, Zack Ellerby.

---

More Details

A Hands-on Tutorial on Apache Spark for Solving Data Science Problems with Fuzzy logic

Organisers:  Isaac Triguero (<isaac.triguero@nottingham.ac.uk>), Alberto Fernández (<alberto@decsai.ugr.es>), Mikel Galar (<mikel.galar@unavarra.es>).

In the era of big data, the leverage of recent advances achieved in distributed technologies enables a novel scenario known as data science, whose main goal is to discover unknown patterns or hidden relations from voluminous data in a faster way. Extracting knowledge from big data becomes a very interesting and challenging task where we must consider new paradigms to develop scalable algorithms. However, computational intelligence models for machine learning, including those that consider fuzzy logic, cannot be straightforwardly adapted to the new space and time requirements. Hence, existing algorithms should be redesigned or new ones developed in order to take advantage of their capabilities in the big data context. Moreover, several issues are posed by real-world complex big data problems besides from computational complexity, and big data mining techniques should be able to deal with challenges such as dimensionality, class-imbalance, and lack of annotated samples among others.

Addressing Big Data becomes a very interesting and challenging task where we must consider new paradigms to develop scalable algorithms. The MapReduce framework, introduced by Google, allows us to carry out the processing of large amounts of information. Its open source implementation, named Hadoop, has allowed the development scalable algorithm becoming de facto standard for addressing Big Data problems. Recently, new alternatives to the standard Hadoop-MapReduce framework have arisen to improve the performance in this scenario, being Apache Spark project the most relevant one. Even working on Spark, the MapReduce framework implies that existing algorithms need to be redesigned or new ones need to be developed in order to take advantage of their capabilities in the big data context.

Data science is a quite recent field of study, and it is still rapidly expanding. In other words, the open directions for novel research are particularly associated with the analysis of the application of fuzzy systems to emerging work scenarios in data science. Some clear examples are data streams, imbalanced classification, or big dimension, among others.

In this practical tutorial, we will first provide a gentle introduction to the problem of Big Data from the perspective of the development of fuzzy-based models. Then, we will dive into the field of Big Data analytics, describing several interesting case studies and real applications on the topic addressed with fuzzy approaches. Finally, we will run a hands-on session on Apache Spark and the Machine Learning library.

Table of contents and estimated duration:

• An Introduction to Big Data Analytics with Fuzzy-based Models (30 minutes)

• Case Studies in Data Science with Fuzzy Approaches (30 minutes)

• Hands-on session (45 minutes)

• Final Discussion (15 minutes)

Intended audience and expected enrollment:

This tutorial is aimed at all those researchers involved in the development of fuzzy models as well as evolutionary algorithms, providing them an overview of the existing technologies to deal with Big Data problems. The audience will also be able to understand the impact of the use of such kind of approaches in Data Science, in particular by means of our most recent research publications on the topic.

This proposal is built up on the success of our previous tutorials at WCCI 2016, CEC 2017, FUZZ-IEEE 2017, WCCI 2018, and FUZZ-IEEE2019. Prof. F. Herrera, Dr. I. Triguero, Dr. Mikel Galar, and Dr. Alberto Fernández delivered these tutorials in their different editions. In all these sessions, an average of 50 participants were involved. Thus, we expect a high number of participants as this topic continues to be an exciting and underexplored field that can have a great impact on the way we analyze and exploit data.

Organiser Bios (Click to expand)Organiser Bios

A Top-Down Approach to Rule-Based Fuzzy Systems

Organiser: Jerry Mendel (<jmmprof@me.com>).

Having recently completed the second edition of a textbook about rule-based fuzzy systems, I have become very concerned about the amount of time and background knowledge that are required before a new person is able to use such systems. Such a large investment of time on the part of a new person, who often needs a solution to a real-world problem within a modest amount of time, may drive that person to use what they think is a competitive methodology that does not require such a large investment. I believe that the major cause for the large amount of time required to learn rule-based fuzzy systems is the widely used bottom-up approach for developing and studying them. This tutorial explains a novel top-down approach to rule-based fuzzy systems, one that begins with the product and then addresses the unique features of the product, without requiring the reader to know anything about fuzzy sets and systems, so that the attendee has something to use with a very small investment of time.
The “products” for rule-based fuzzy systems are input-output equations (formulas). What a new end-user needs to know for each formula is: (1) a simple explanation for it, (2) what the unique features are for it, (3) how it can be used to solve real-world problems, (4) how it competes with formulas from other technologies that might also be used for the same real-world applications; and, (5) where a computer program can be found for it. This tutorial will illustrate the top-down approach beginning with simple products and concluding with advanced products and will address what an end-user needs to know in order to use it, i.e. items (1)–(4) just stated. Learning about rule-based fuzzy systems can be greatly compressed by using the top-down approach that is described in this Tutorial. It will let you cover them in an existing course in a small number of lectures.

Organiser Bio (Click to expand)Organiser Bio

Besides fuzzy sets: new theories for promptly moving your crisp systems’ full power to the fuzzy side.

Organiser: Daniel Sanchez (<daniel@decsai.ugr.es>).

The first objective of the tutorial is to introduce the theory of Representation by Levels (RLs) to
the audience, as well as closely related theories such as Dubois and Prade’s Gradual Sets, and
Trevor Martin’s X-μ approach. These theories allow us to represent “fuzziness” in subsets of X by means of functions ranging from (0,1] to the power set of X, instead of the usual fuzzy set
approach, where membership functions range from X to [0,1]. Representations by levels can be used in order to represent “fuzziness” in other mathematical objects other than sets, like elements. For instance, real numbers affected by fuzziness are represented by functions ranging from (0,1] to the set of real numbers.

Representations by levels can be employed as a way to “fuzzify” any crisp system, algorithm,
concept or theory in a single, unique and straightforward way, and keeping all the properties of the original crisp counterpart. This is due to i) operations being performed in each level
independently, and ii) no restriction being required among levels (for instance, no nesting is
required in the representation for sets, contrary to alpha-cut representations of fuzzy sets). Such features cannot be provided by fuzzy set theories and their extensions; paradigmatic examples are the unability of fuzzy set theories to keep all the Boolean properties of set operations simultaneously, and the unability of fuzzy numbers to keep the algebraic structure of crisp numbers. This way, representation by levels allows to promptly adapt crisp systems to deal with fuzzy information without the necessity to determine the most appropriate operators to extend set operations, arithmetic operations, or any other kind of operator or procedure. This easyness, together with the unique feature that the resulting fuzzy system keeps all the properties of the crisp version, make Representations by Levels adequate for incorporating fuzzy sets in the development of systems that i) need to deal with fuzzy information (concepts, data, etc.), ii) need to keep properties that cannot be provided simultaneously by fuzzy sets, and iii) require the availability of promptly developed prototypes.

Though different in structure, operations and properties, RLs and fuzzy sets are tightly coupled and complementary. They are tightly coupled because i) the collection of alpha-cuts of a fuzzy set with a finite number of membership degrees (which is the case in practice for both
computers and humans) is a particular case of RLs, and ii) fuzzy sets can be viewed as probability measures defined on RLs (different RLs providing the same fuzzy set when measuring, with degrees of individual elements meaning “probability of membership to a level taken at random”). But they are also complementary in the sense that fuzzy sets are much easier to understand for users than RLs. These ideas lead to the notion of RL-systems: systems in which the input and output take the form of fuzzy sets, and where the internal representation and operations, hidden to the users, is performed by means of RLs. Remarkably i) the output fuzzy sets are not employed as input of another system, the RLs being used instead, and ii) the output fuzzy sets do not correspond to those that can be obtained by ordinary fuzzy set theories and extensions in general, as the latter do not keep the crisp properties of the original system in general.

Overall, the tutorial is intended to show the attendees tools and methodologies about how to
extend crisp systems to the fuzzy case as RL-systems, allowing for a prompt adaptation able to keep the properties of the crisp case, and based on fuzzy sets as input and output of the  system, hence improving understandability and taking benefit of the available techniques for
communicating with users using fuzzy set theory. This way, RLs are not conceived as a substitute or a better theory than fuzzy set theory, but a complement which enlarge our available toolbox for the development of fuzzy systems, concepts, algorithms and theories with new powerful capabilities, following the idea of RL-systems.

The tutorial will take the form of a presentation of the concepts as stated above, with theoretical background and both theoretical and practical examples illustrating the theory of
Representations by Levels and its application to specific problems like cardinality, probability
and quantification, arithmetics of RL-numbers, clustering, association rule mining and referring expression generation within Natural Language Generation, among others.

Outline of the covered material:
A tentative outline is the following:
1. Representations by levels: basic concepts and operations. Related theories.
2. From fuzzy sets to RLs and back: a non-biyective relation.
3. RL-systems: how to promptly take your crisp system to the fuzzy side.
4. Application examples.
5. Open challenges.

Though the first publications on the topic date back to 2008, RLs are not yet widely known by the fuzzy set community (not to mention outside). However, many papers with theoretical results and practical applications have been published since then, showing the suitability of the approach. In our opinion, RLs also have the potential to push fuzzy set theory into other scientific communities like those of Natural Language Generation or Artificial Intelligence in general. Particularly, in these communities there are several authors that point to a drawback of fuzzyset theories that prevent them from considering it in practice: the lack of clear rules for choosing the most suitable operators for each problem, and the unability to keep the ordinary properties of the crisp case. RL-systems are able to provide both things while allowing for partial membership both in the input/output by means of fuzzy sets, and in the internal representation and operations via RLs. In our conversations with some some relevant authors in the abovementioned communities, they have found the approach interesting and worth exploring.  On the other hand, the fuzzy set community is nowadays highly interested in promoting fuzzy sets in the AI community, and we envisage RLs as a way to offer a novel “fuzzy” tool able to possibly overcome existing reluctances to consider proposals coming from our area. Overall, we think that the timeliness of our proposal is justified and, if accepted, our intention is to disseminate the call for participation both within the fuzzy set and other communities such as those working in AI and NLG.

Proposed duration is one hour and a half. The proposer is one of the co-authors of the RL theory, and has been working and publishing results in the field since 2008, including collaborations with Didier Dubois, co-author of the closely related Gradual Sets. This collaboration led to a joint paper in which RLs are employed for developing a new fuzzy clustering technique, remarkably able to solve the problem posed by Bezdek and Harris in 1978 about how to obtain convex decompositions of similarity relations for clustering.

Organiser Bio (Click to expand)Organiser Bio

Building interpretable fuzzy inference systems in Python

Organisers: Uzay Kaymak  (<u.kaymak@ieee.org>), Marco Nobile (<m.s.nobile@tue.nl>), Simone Spolaor (<m.s.nobile@tue.nl>), Caro Fuchs (<c.e.m.fuchs@tue.nl>).

One definition of artificial intelligence (AI) is “building systems that act rationally”, an approach that represented the driving force of AI research in the last decade (Russel & Norvig, 2000). However, the practical deployment of AI-based decision systems is hampered in all domains in which liability plays a major role (e.g., health-care). As a matter of fact, regardless of the performances of a system, the decisions that an AI suggests do not generate trust as long as their rationale cannot be inspected and understood by domain experts. Stated otherwise, the development of interpretable systems represents an enabling technology for widespread adoption of AI. It is also worth noting that even the EU General Data Protection Regulation explicitly mandates the “right of explanation” (see, e.g., Recital 71 of EU GDPR), so that no black-box decision system should be used to make automatic decisions, having legal effects on a subject, if they cannot be explained.
This tutorial is designed to describe how interpretable AI systems can be easily created by using fuzzy systems, either employing domain knowledge or according to a data-driven approach. At the end of the tutorial, attendees will be able to create interpretable AI systems in the form of fuzzy models, by leveraging two recently published Python libraries, namely Simpful and pyFUME. Attendees will be guided in the creation of such models and the evaluation of their results. In order to do so, they will be provided with some usage examples based on bio-medical use cases and invited to apply the presented methods to their own data sets.

The session can be subdivided into three main Sections:
1. theoretical background (15 mins + 5 mins Q&A);
2. knowledge-driven modeling (20 mins + 10 mins Q&A)
3. data-driven modeling (40 mins + 10 mins Q&A)
In Section #1, we will explain the motivation for the development of interpretable AI systems. Then, we will explain the theoretical foundations of our approach based on Fuzzy Logic.
Section #2 will be focused on the implementation according to domain knowledge. The library exploited for this part will be Simpful. Simpful’s syntax and API will be described in detail.
In Section #3, will be focused on the implementation according to domain knowledge. The library exploited for this part will be pyFUME. The functioning and the API of pyFUME will be described in detail.

The tutorial assumes knowledge of the Python 3 programming language. Explanations for the setup of the system (i.e., how to install and test Simpful and pyFUME) can be found at the following web address: bit.ly/tutorial-fuzzieee.

At the end of the tutorial, the attendees will be able to autonomously build up an interpretable AI system, either based on domain knowledge or automatically inferred from available data.

This tutorial will cover the motivations and the aims behind the choice of building interpretable AI systems. In particular, the tutorial will focus on the use of fuzzy systems as a tool to achieve more interpretable models in decision making and complex systems simulation. In order to facilitate aspiring scientists and newcomers to the field, a brief introduction to the fundamentals of fuzzy inference systems will be provided, including an outline of fuzzy set theory and fuzzy reasoning and logic.

The practical part of the tutorial will include the presentation of the two Python libraries (i.e., Simpful and pyFUME). In the Simpful session, attendees will learn how to define fuzzy inference systems within this novel library. Once attendees will be familiar with Simpful, the tutorial will move on to the next session involving data-driven approaches. In this session, attendees will learn to infer (minimal) fuzzy models from given data sets by exploiting pyFUME.

Potential audience: Both Simpful and pyFUME have been developed to fit the needs of academics and practitioners. They are both easy to use (because of their pre-implemented and user-friendly pipelines) and flexible for scientists that wish to fine-tune each step of the fuzzy model estimation process. We therefore envision that this tutorial will be useful to aspiring scientists and industrial partners, who wish to learn more about fuzzy models and their development, but also to senior scientists who search for a tool to help them conduct their studies. They will also be able to assess first-hand the advantages the existing libraries offer for their applications.

Plans for promotion: ​ Mailing lists (EUSFLAT, SIKS, IFSA, IEEE CIS Newsletter), ads in the context of the Summer School on Fuzzy Logic and Applications (including SFLA Facebook and Twitter page) and in the events organized by the IEEE CIS task force on advanced representation in biological and medical search and optimization.

Timeliness of the given tutorial: In the past few years, Artificial Intelligence (AI) has emerged as an important and far-reaching research topic with applications in many fields. This development has led to the recent trend to pursue interpretable or explainable AI (XAI), which is not only believed to build trust between the AI system and its user but can also give insights into the way decisions are taken and therefore create new knowledge.

In this tutorial, we show the participants how they can use their own data to create interpretable AI systems in the form of fuzzy models leveraging Simpful and pyFUME. Both Simpful and pyFUME were implemented in Python 3, one of the most popular and widespread programming languages. Thanks to its high readability and expressive power, it found application in several fields which deal with computer and data science, as well as in multidisciplinary research and industrial sectors. Both libraries were published in peer-reviewed journals during the last year and are available in public repositories. Participants will be provided with updated information and practical examples on the usage of the libraries. They will also be invited to employ the libraries on their own datasets and use-cases.

Duration: ​ 100 minutes (divided in 3 blocks with Q&As)

Organiser Bios (Click to expand)Organiser Bios

Explainable Fuzzy Systems: Paving the way from Interpretable Fuzzy Systems to Explainable Artificial Intelligence

Organisers: Jose Alonso (<josemaria.alonso.moral@usc.es>), Ciro Castiello (<ciro.castiello@uniba.it>), Corrado Mencar (<corrado.mencar@uniba.it>), Luis Magdalena (<luis.magdalena@upm.es>).

In the era of the Internet of Things and Big Data, data scientists are required to extract valuable knowledge from the given data. They first analyze, cure and pre-process data. Then, they apply Artificial Intelligence (AI) techniques to automatically extract knowledge from data. Actually, AI has been identified as the “most strategic technology of the 21st century” and is already part of our everyday life [1]. The European Commission states that “EU must therefore ensure that AI is developed and applied in an appropriate framework which promotes innovation and respects the Union’s values and fundamental rights as well as ethical principles such as accountability and transparency”. It emphasizes the importance of eXplainable AI (XAI in short), in order to develop an AI coherent with European values: “to further strengthen trust, people also need to understand how the technology works, hence the importance of research into the explainability of AI systems”. Moreover, as remarked in the last challenge stated by the USA Defense Advanced Research Projects Agency (DARPA), “even though current AI systems offer many benefits in many applications, their effectiveness is limited by a lack of explanation ability when interacting with humans”. Accordingly, users without a strong background on AI, require a new generation of XAI systems. They are expected to naturally interact with humans, thus providing comprehensible explanations of decisions automatically made.

XAI is an endeavor to evolve AI methodologies and technology by focusing on the development of agents capable of both generating decisions that a human could understand in a given context, and explicitly explaining such decisions. This way, it is possible to verify if automated decisions are made on the basis of accepted rules and principles, so that decisions can be trusted and their impact justified.
Even though XAI systems are likely to make their impact felt in the near future, there is a lack of experts to develop the fundamentals of XAI, i.e., ready to develop and to maintain the new generation of AI systems that are expected to surround us soon. This is mainly due to the inherent multidisciplinary character of this field of research, with XAI researchers coming
from heterogeneous research fields. Moreover, it is hard to find XAI experts with a holistic view as well as wide and solid background regarding all the related topics.
Consequently, the main goal of this tutorial is to provide attendees with a holistic view of fundamentals and current research trends in the XAI field, paying special attention to fuzzy-grounded knowledge representation and how to enhance human-machine interaction.

Plan
The tutorial will cover the main theoretical concepts of the topic, as well as examples and real applications of XAI techniques. In addition, ethical and legal aspects concerning XAI will also be considered. 
The tutorial duration is about 2 hours, with four blocks of 30min each covering the different questions under consideration:

• We will first introduce the general ideas behind XAI by referring to real-world problems that would take great benefit from XAI technologies. Also, some of the most recent governmental and social initiatives which favor the introduction of XAI solutions in industry, professional activities and private lives will be highlighted. This part is therefore devoted to motivating the audience on the potential impact of XAI in everyday life and, therefore, on the importance of its scientific and technical investigation.

• The second part of the tutorial will be devoted to a gentle introduction of the main methods for XAI at the state of the art. This part is cross-field, but it globally falls within the realm of Computational Intelligence. The idea of “opening the black box” will be stressed (being the “black-boxes” models designed through Machine Learning techniques, such as deep neural networks), as well as several approaches for dealing with the concept of “explanation”. Special attention will be focused on natural language explanation (such as, natural language generation of explanations, argumentative techniques, human-machine interactions, etc.).

• The third part of the tutorial will take into account the special role of fuzzy logic in XAI. It will be shown that fuzzy logic offers special features that enable a rich representation of concepts expressible in natural language, therefore it might be a privileged choice for explanation generation and processing. Interpretable fuzzy systems use fuzzy logic to represent knowledge that is easy to be read and understood by users: hence, they will be revisited from the point of view of XAI and natural language generation.

• Space will be reserved to address some currently open research directions, such as the evaluation of XAI systems, to stimulate the audience, especially young researchers, to explore and contribute to this new field. Finally, the tutorial will offer also the opportunity to present software tools implementing some XAI technologies. Some of these tools will be used to show a use case in order to make XAI results tangible to the audience.

Intended Audience
This tutorial is of interest for researchers, practitioners and students (PhD or Master students) working in the fields of Artificial and Computational Intelligence; with special emphasis on Fuzzy Logic. Since our aim is to provide attendees with a holistic view of fundamentals and current research trends in the XAI field, and having in mind the broad interest of the topic, the presentations will be designed to be accessible to everyone no matter their background.

Outline of the Covered Material

1. Introduction: Motivating Principles and Definitions.
2. Review of the most Outstanding Approaches for Designing and Developing XAI systems
   1. Review of the most outstanding approaches for Opening Black Boxes
   2. Review of Natural Language Technology for XAI
   3. Review of Argumentation Technology for XAI
   4. Review of Interactive Technology for XAI
   5. Review of Software Tools for XAI
3. Fuzzy Technology for XAI
   1. Building Interpretable Fuzzy Systems
   2. Building Explainable Fuzzy Systems
   3. Software Tools
4. Evaluation of XAI systems
5. Use Case
6. Concluding Remarks

Organiser Bios (Click to expand)Organiser Bios

Fuzzy Systems for Brain Sciences & Interfaces.

Organisers: Javier Andreu-Perez (<javier.andreu@essex.ac.uk>), Ching Teng Lin (<chin-teng.lin@uts.edu.au>).

In this tutorial will be introduced the latest work in the involvement of fuzzy systems in neuroscience and neuro-engineering. Given the important challenges associated with the processing of brain signals obtained from neuroimaging modalities, fuzzy sets and systems have been proposed as a useful and effective framework for the analysis of brain activity as well as to enable a direct communication pathway between the brain and external devices (brain computer/machine interfaces). While there has been increasing interest in these questions, the contribution of fuzzy logic sets, and systems has been diverse depending on the area of application in neuroscience. On the one hand, considering the decoding of brain activity fuzzy sets and systems represent an excellent tool to overcome the challenge of processing extremely noisy signals that are very likely to be affected by high uncertainty. On the other hand, as regards neuroscience research, fuzziness has equally been employed for the measurement of smooth integration between synapses, neurons, and brain regions or areas. In this context, the proposed tutorial aims at introducing the latest advancements and enabling a discussing among researchers interested in employing fuzzy sets, logic and systems for the analysis of brain signals and neuroimaging data, including related disciplines such as computational neuroscience, brain computer/machine interfaces, neuroscience, neuroinformatics, neuroergonomics, affective neuroscience, neurobiology, brain mapping, neuroengineering, and neurotechnology. In this tutorial we will explain the current-state of the art of the application of fuzzy systems in the domain of neuroscience. Moreover, we will also go explain their potential applications in the field of Brain-Computer Interfaces, and their performance with respect to competing methods.

Outline of the covered material

In this tutorial is intended to provide to the attendants the needed knowledge in the field of neuroscience and of fuzzy systems to be able to produce work in the field. The tutorial will not require the attendants to have a prior knowledge on either neuroscience or fuzzy systems. The format of the session will cover the following syllabus:

1. Introduction about Non-Invasive Neuroimaging methods (EEG & fNIRS)
2. Rationale about Fuzzy Systems in Neuroscience.
3. Introduction to methodologies for Analysis of Brain Signals in Neuroscience (connectomes, brain activation analysis, ROI analysis, brain decoding).
4. Introduction to the mechanics of helpful structures in Fuzzy Logic and Systems to use with Brain data (Type-1 Sets Extensions, Neuro-Fuzzy, Evolutionary Fuzzy Models, fuzzy integrals and fuzzy entropies).
5. Current advanced methods of Fuzzy Systems for Neuroscience Applications.
6. Current advanced methods of Fuzzy Systems for Brain Computer Interfaces.
7. Publicly available software and toolboxes.

Plan of the session:
The session is planned to last 90 minutes and will consist of three main sessions (chapters).

1. First Session (45 minutes): Covering points 1-4 (introduction)
2. Second Session (30 minutes): Covering points 5 – 6.
3. Third Session (15 minutes): Covering point 7.

Justification:
The rationale behind this tutorial is very timely. Current methodologies for the analysis of brain signals used in Neuroscience are very basic and sometime over-simplistic. The use of Machine Learning has just timidly started to be implemented in the field. Usually, very simple statistical methods are used to analyse the highly chaotic and noisy signals arising from neuroimaging. Fuzzy Systems have brought a better performance into many fields and to model complex systems. In neuroscience, they can furnish investigators with a new set of
the research methods that will help them to maximise the informational power of non-invasive neuroimaging techniques such as EEG and fNIRS. In comparison with other highly expensive neuroimaging modalities such as fMRI or DTI, non-invasive techniques of neuroimaging can be used in more naturalistic scenarios to study the brain at work. Also, they can power neuro-prothesis or help in rehabilitation tasks for brain injury. The issue is that such non-invasive techniques provides a vision of the neural processes that is less precise than their expensive counterparts. The application of fuzzy systems in this domain is helping to get out of the most of these non-invasive neuro-imaging modalities, and in this tutorial, we will introduce attendants with all major developments to date.

Organiser Bios (Click to expand)Organiser Bios

How big is too big? Clustering in (static) BIG DATA with the Fantastic 4

Organiser: James Bezdek (<jcbezdek@gmail.com>).

For this talk “big” refers to the number of samples (N) and/or number of dimensions ( P) in static sets of feature vector data; or the size of (NxN) (similarity or distance) matrices for relational clustering. Objectives of clustering in static sets of big numerical data are acceleration for loadable data and approximation for non-loadable data. The Fantastic Four are four basic (aka “naïve”) classical clustering methods:

Gaussian Mixture Decomposition (GMD, 1898)
Hard c-means (often called “k-means,” HCM, 1956)
Fuzzy c-means (reduces to hard k-means in the limit, FCM, 1973)
SAHN Clustering (principally single linkage (SL, 1909))

This talk describes approximation of literal clusters in non-loadable static data. The method is sampling followed by very fast (usually 1-2% of the overall processing time) non-iterative extension to the remainder of the data with the nearest prototype rule. Three methods of sampling are covered: random, progressive, and MaxiMin. The first three models apply to feature vector data and find partitions by approximately optimizing objective function models with alternating optimization (known as expectation-maximization (EM) for GMD). Numerical examples using various synthetic and real data sets (big but loadable) compare this approach to incremental methods (spH/FCM and olH/FCM) that process data chunks sequentially.

The SAHN models are deterministic, and operate in a very different way. Clustering in big relational data by sampling and non-iterative extension begins with visual assessment of clustering tendency (VAT/iVAT). Extension of iVAT to scalable iVAT (siVAT) for arbitrarily large square data is done with Maximin sampling, and affords a means for visually estimating the number of clusters in the literal MST of the sample. siVAT then marries quite naturally to single linkage (SL), resulting in two offspring: (exact) scalable SL in a special case; and clusiVAT for the more general case. Time and accuracy comparisons of clusiVAT are made to crisp versions of three HCM models; HCM (k-means), spHCM and olHCM; and to CURE. Experiments synthetic data sets of Gaussian clusters, and various real world (big, but loadable) are presented.

Organiser Bio (Click to expand)Organiser Bio

New Optimization Techniques for TSK Fuzzy Systems: Classification and Regression

Organiser: Dongrui Wu (<drwu@hust.edu.cn>).

Description of the goal of the proposed tutorial (max. 4000 characters):
Takagi-Sugeno-Kang (TSK) fuzzy systems have achieved great success in numerous applications, including both classification and regression problems. Many optimization
approaches have been proposed for them.

There are generally three strategies for fine-tuning the TSK fuzzy system parameters after initialization: 1) evolutionary algorithms; 2) gradient descent (GD) based algorithms; and, 3) GD plus least squares estimation (LSE), represented by the popular adaptive-network-based fuzzy inference system (ANFIS). However, these approaches may have challenges when the size and/or the dimensionality of the data increase. Evolutionary algorithms need to keep a large population of candidate solutions, and evaluate the fitness of each, which result in high computational cost and heavy memory requirement for big data. Traditional GD needs to compute the gradients from the entire dataset to iteratively update the model parameters, which may be very slow, or even impossible, when the data size is very large. The memory requirement and computational cost of LSE also increase rapidly when the data size and/or dimensionality increase. Additionally, our research demonstrated that ANFIS may result in significant overfitting in regression problems.

This tutorial explains the functional equivalence between TSK fuzzy systems and certain neural networks, mixture of experts, CART and stacking ensembles; hence, optimization techniques for the latter may be used for TSK fuzzy systems. It then introduces some newly proposed novel techniques for mini-batch gradient descent (MBGD) based optimization of TSK fuzzy systems, for both classification and regression. These techniques were mainly extended from the training of deep neural networks and mixture-of-experts models. The proposed algorithms can be applied to datasets of any size, but are particularly useful for big data applications. Their source code is also available online. We believe this tutorial will further facilitate the applications of TSK fuzzy systems.

The session mainly consists of three parts:
1. Functional Equivalence between TSK Fuzzy Systems and Neural Networks, Mixture of Experts, CART and Stacking Ensemble Learning
Fuzzy systems have achieved great success in numerous applications. However, there are still many challenges in designing an optimal fuzzy system, e.g., how to efficiently optimize its parameters, how to balance the trade-off between cooperation and competitions among the rules, how to overcome the curse of dimensionality, how to increase its generalization ability, etc. Literature has shown that by making appropriate connections between fuzzy systems and other machine learning approaches, good practices from other domains may be used to improve the fuzzy systems, and vice versa. This part gives an overview on the functional equivalence between TSK fuzzy systems and four classic machine learning approaches — neural networks, mixture of experts, classification and regression trees (CART), and stacking ensemble regression — for regression problems. We also point out some promising new research directions, inspired by the functional equivalence, that could lead to solutions to the aforementioned problems. To our knowledge, this is so far the most comprehensive overview on the connections between fuzzy systems and other popular machine learning approaches, and hopefully will stimulate more hybridization between different machine learning algorithms.
2. New Optimization Technique for TSK Fuzzy Classifiers: Mini-Batch Gradient Descent, Uniform Regularization, and Batch Normalization
TSK fuzzy systems are flexible and interpretable machine learning models; however, they may not be easily optimized when the data size is large, and/or the data dimensionality is high. This part proposes a MBGD based algorithm to efficiently and effectively train TSK fuzzy classifiers. It integrates two novel techniques: 1) uniform regularization (UR), which forces the rules to have similar average contributions to the output, and hence to increase the generalization performance of the TSK classifier; and, 2) batch normalization (BN), which extends BN from deep neural networks to TSK fuzzy classifiers to expedite the convergence and improve the generalization performance. Experiments on 12 UCI datasets from various application domains, with varying size and dimensionality, demonstrated that UR and BN are effective individually, and integrating them can further improve the classification performance.
3. New Optimization Technique for TSK Fuzzy Regression Models: Mini-Batch Gradient Descent, Regularization, AdaBound, and DropRule
TSK fuzzy systems are very useful machine learning models for regression problems. However, to our knowledge, there has not existed an efficient and effective training algorithm that ensures their generalization performance, and also enables them to deal with big data. Inspired by the connections between TSK fuzzy systems and neural networks, we extend three powerful neural network optimization techniques, i.e., MBGD,regularization, and AdaBound, to TSK fuzzy systems, and also propose three novel techniques (DropRule, DropMF, and DropMembership) specifically for training TSK fuzzy systems. Our final algorithm, MBGD with regularization, DropRule and AdaBound (MBGD-RDA), can achieve fast convergence in training TSK fuzzy systems, and also superior generalization performance in testing. It can be used for training TSK fuzzy systems on datasets of any size; however, it is particularly useful for big datasets, on which currently no other efficient training algorithms exist.

This session will be mainly powerPoint based tutorial. We will also show how the proposed
algorithms work in Matlab.
Outline of the covered material:
1. Introduction to TSK Fuzzy Systems
2. Functional Equivalence between TSK Fuzzy Systems and Neural Networks, Mixture of
Experts, CART and Stacking Ensemble Regression
3. New Optimization Technique for TSK Fuzzy Classifiers: Mini-Batch Gradient Descent,
Uniform Regularization, and Batch Normalization
4. New Optimization Technique for TSK Fuzzy Regression Models: Mini-Batch Gradient
Descent, Regularization, AdaBound, and DropRule
5. Concluding Remarks Justification, which should include the potential audience and the timeliness of the given tutorial, proposed duration, as well as the qualifications of the proposer(s):
Potential audience: Researchers, practitioners, and graduate students in fuzzy systems. 40
audiences are expected.
Timeliness: The tutorial is based on the following three latest research works.
1. D. Wu, C-T Lin, J. Huang* and Z. Zeng*, “On the Functional Equivalence of TSK Fuzzy Systems to Neural Networks, Mixture of Experts, CART, and Stacking Ensemble Regression,” IEEE Trans. on Fuzzy Systems, 28(10):2570-2580, 2020.
2. Y. Cui, D. Wu* and J. Huang*, “Optimize TSK Fuzzy Systems for Classification Problems: Mini-Batch Gradient Descent with Uniform Regularization and Batch Normalization,” IEEE Trans. on Fuzzy Systems, 2020, accepted.
3. D. Wu*, Y. Yuan, J. Huang and Y. Tan*, “Optimize TSK Fuzzy Systems for Regression Problems: Mini-Batch Gradient Descent with Regularization, DropRule and AdaBound
(MBGD-RDA),” IEEE Trans. on Fuzzy Systems, 28(5):1003-1015, 2020.

Organiser Bio (Click to expand)Organiser Bio

Recent Developments in Artificial Intelligence and Fuzzy Systems

Organisers: Alexander Gegov (<alexander.gegov@port.ac.uk>), Uzay Kaymak (<u.kaymak@ieee.org>), Joao Sousa (<jmsousa@tecnico.ulisboa.pt>).

Tutorial goalThe tutorial highlights recent and current developments in AI and Fuzzy Systems. It has an educational focus that is complemented by latest research results in the field. The tutorial introduces basic concepts that are illustrated with a range of examples and case studies from different application areas.

Tutorial planThe tutorial includes two parts. The first part focuses on general aspects of AI and Fuzzy Systems. The second part focuses on specific aspects of AI and Fuzzy Systems. Both parts include discussion with the audience. The expected size of the audience is about 50 or more.

Tutorial outlineThe first part of the tutorial introduces recent developments, rational justification, current success, subject areas and wider popularity of AI. It highlights Human Intelligence as a role model for AI and Computational Intelligence as a driving force behind AI. Computational Intelligence techniques such as Fuzzy Systems, Neural Networks and Evolutionary Algorithms are discussed in the context of their ability to imitate different aspects of Human Intelligence in AI. Approaches such as Expert Systems and Machine Learning are also discussed in the context of their ability to work in a complementary way with knowledge and data.
The second part of the tutorial introduces Cybernetics and Complex Systems as research areas that are closely related and complementary to AI. It underlines the different aspects of system complexity and their impact on the performance of AI based models. Current hot topics such as Big Data, Deep Learning and Deep Fuzzy Systems are discussed in the context of feature extraction and selection as well as structural and parametric identification. Other AI related aspects such as current limitations, future developments, main challenges, application areas, case studies and performance evaluation are also discussed.

Tutorial justificationThe potential audience for the tutorial includes undergraduate students, postgraduate students, researchers, academics and practitioners. The proposed duration of the tutorial is 1.5 hours. The organisers are planning to set up a website to promote the tutorial. The tutorial topic is quite timely in the proposed educational and research setting in view of the current popularity of AI and its rising impact on everyday life. The tutorial proposers are established researchers in the field with many years of experience and involvement as Technical Committee Members for IEEE Societies and Associate Editors for IEEE Transactions on Fuzzy Systems.

Organiser Bios (Click to expand)Organiser Bios

Towards Trusting AI/ML Decision-Making of Autonomous Systems <Cancelled>

Organiser: Stanton Price (<stanton.r.price@usace.army.mil>)

Autonomous systems have a growing presence throughout numerous industries; from self-driving cars to warehouse distribution robots, surgical robots, and defense systems. As these autonomous systems begin to make decisions of increasing complexity while simultaneously having potentially life-impacting consequences, it becomes critical to trust the decisions being made. Further, situations in which unexpected outcomes arise (good and/or bad), there is a need to be able to explain/understand why the machine performed the action(s) that led up to the unforeseen consequence. One way to get closer to obtaining trust in these systems is through understanding why decisions are made. This tutorial will cover the topic of gaining user trust in autonomous systems through understanding AI/ML decision-making process and how this area is uniquely primed for fuzzy set theory. Additionally, this tutorial will take a hybrid approach, mixing interactive learning experiences with a ‘reverse’ problem-owner-led tutorial seeking academic engagement to foster potential collaboration.

Tutorial Outline
Introduction to autonomous systems (more heavily focused on autonomous ground
vehicles) (5-10 minutes) Overview of fuzzy sets (5-10 minutes) Overview of numerous eXplainable AI methodologies (5-10 minutes) Demo (20-30 minutes): Deep dive of the demo
Identify task-specific goals (e.g., identify obstacles that could impede autonomous ground vehicle mobility)

• Identify holes/gaps in current techniques

•  Identify areas where fuzzy set theory could enhance understanding of the decision-making process

Deep dive into identified holes/gaps and how fuzzy set theory could play a key role
in addressing the identified shortcomings (10-15 minutes) ‘Reverse’ the tutorial, probe for academic engagement and collaboration (15-30 minutes)

Justification
Given the relevancy of autonomous systems, mixed with the growing interest in AI/ML explainability of these systems, this tutorial is expected to draw an appreciable audience.
Further, the audience is expected to expand across industry, government, and academia. This tutorial will be promoted throughout various U.S. DoD agencies and Research Centers (e.g., Army, Navy, ERDC, etc.) and through academic collaborators (both U.S. and international). As detailed above in the Outline, this tutorial is expected to last approximately 1.5 hours. Dr. Price leads multiple projects that include aiding the development of state-of-the-art autonomous systems as well as developing state-of-the-art XAI methodologies for various applications. He is well versed in these technologies and fuzzy set theory, providing insights into how fuzzy sets could play a critical role in understanding autonomous systems’ decision-making process.

Organiser Bio (Click to expand)Organiser Bio

Uncertainty modeling in adversarial learning

Organiser: Xi-Zhao Wang (<xizhaowang@ieee.org>).

This tutorial with the title Uncertainty modeling in adversarial learning is aiming at providing participants with some fundamental concepts of uncertainty such as fuzziness and ambiguity, some theoretical analysis and key techniques on uncertainty modeling, and their applications to improve the adversarial robustness of deep neural networks.

Specifically, the tutorial goal is
(1) To provide students or young scholars (who just entered this area) with a basic course of machine learning under uncertainty environment, focusing on uncertainty modeling and its impact in the whole process of learning on the performance of learning algorithms;
(2) To give an introductory course for participants who are working on the recently popular topic, adversarial learning, especially for these who are working on adversarial learning of some specific domains and want to use the uncertainty modeling tool to improve the adversarial robustness of their systems;
(3) To offer some useful guidelines for ongoing PhD students who are devoted to study their PhD projects to model and solve some key issues of uncertainty modeling in adversarial learning; and
(4) To talk with participants who are specialists or senior scholars about some new treads in the area of learning from fuzzy examples and adversarial learning in which uncertainty modeling plays a key role for the improvement of adversarial robustness.

The tutorial will contain the following content.
(1) Introduction to fuzziness and uncertainty. Uncertainty is a natural phenomenon in machine learning, which can be embedded in the entire process of data preprocessing, learning and reasoning. For example, the training samples are usually imprecise, incomplete or noisy, the classification boundaries of samples may be fuzzy, and the knowledge used for learning the target concept may be rough. Uncertainty can be used for selecting extended attributes and informative samples in inductive learning and active learning respectively. If the uncertainty can be effectively modeled and handled during the process of processing and implementation, machine learning algorithms will be more flexible and more efficient. This part will focus on the uncertainty definition and relationships/differences among different uncertainties.
(2) Adversarial examples. An adversarial example is an instance with intentional but small feature  perturbations that cause a machine learning model to make a false prediction. Usually, we call this process of generating perturbations as adversarial attack, and the ability of model to resist adversarial attack as adversarial robustness. Adversarial examples have become a bigger-than-ever problem with the quick developments of deep learning. Nowadays, applications of deep learning combined with real-world scenarios are becoming wider and deeper, for example, deep learning has been applied to autonomous driving. We have many adversarial cases in real life, for example, a self-driving car crashes into another car because it ignores a stop sign; someone places a picture over the sign, which looks like a stop sign with a little dirt for humans but is designed to resemble a parking prohibition sign for the car’s software recognition. It is undoubtedly very dangerous. This part will concentrate on the adversarial example challenges and their current handling strategies.
(3) Prediction uncertainty in adversarial robustness. This is to show that the representation, measure, and handling of uncertainty have a significant impact on the performance of adversarial robustness. Recent research shows that, through increasing model prediction uncertainty when training the model, the classification boundary surface of the model will be more balanced. It leads to the distance between boundary and data as far aspossible, which significantly increase the difficulty of attacker based on feature perturbation. This part will talk mainly about the relationship between prediction uncertainty and adversarial robustness.
(4) Analyzing the uncertainty of parameters in convolutional layers. Recent researches mainly focus on uncertainty of data and prediction by measuring the fuzziness, entropy or mutual information to improve adversarial robustness of model. However, in this part, we will introduce the uncertainty of parameters in convolutional layers of a deep neural network for attack to improve the adversarial robustness of model. Additionally, this part will introduce a Min-Max property existing in parameters of convolution process and theoretically show its validation. The parameters of convolutional layers would be interpreted from the perspective of uncertainty.
(5) Concluding remarks. This part will give an overview on learning with uncertainty from adversarial examples, summarizing the results acquired in recent study on adversarial learning problems.

Outline of the covered material:
The tutorial outline with brief elaboration is listed below.
1. A brief introduction to adversarial learning. This section introduces the existing definitions of adversarial examples, the explanation of existence of adversarial examples and the application of adversarial learning in various areas such classification, objective detection and semantic recognition etc. (13PPTs)
2. General challenge of adversarial learning. It includes explanation of existence of adversarial examples, constructing most powerful attack, constructing certified defense, obfuscated gradient, and accuracy- robustness. (15PPTs)
3. Current adversarial attacks and adversarial defenses. It covers white-box attacks, black-box attacks, distillation training, adversarial training and various normalization. (10PPTs)
4. Uncertainty. It covers definitions of several specific uncertainties including entropy, fuzziness, ambiguity, and more; the difference and similarity among several uncertainties; criterion of uncertainty modeling, representation, measure, and processing; impact of uncertainty processing on adversarial learning; what role it plays?; how does it affect adversarial learning?; and why does it work? (25PPTs)
5. Uncertainty in adversarial learning. This part includes (1) the relationship between the prediction uncertainty and adversarial robustness; (2) uncertainty of parameters in convolutional layers of a deep neural network for attack. (20PPTs)
6. Concluding remarks. It gives some comments on how to modeling and handling uncertainty of the model to improve the adversarial robustness of a deep neural network with convolution layers. (3PPTs)

Justification, which should include the potential audience and plans for promotion, the timeliness of the given tutorial, proposed duration, as well as the qualifications of the proposer(s): The potential audience includes (1) students or young scholars who want to have a course of adversarial learning with uncertainty modeling; (2) participants who are working on adversarial examples and want to use the uncertainty
modeling tool to improve their learning system performance; (3) ongoing PhD students whose PhD project if related to learning with uncertainty; and (4) specialists or senior scholars who want briefly know some new trends in uncertainty modeling in adversarial learning.
The total number of participants is expected to be 100 more or less. In schedule, the tutorial id delivered in 2 hours with a 20-minute break. Specifically, based on the contend table in previous sections, the time is allocated as: A brief introduction to adversarial learning (15 minutes), General challenge of adversarial learning (10 minutes), Current strategy of adversarial attack and defense (10 minutes), Uncertainty (25 minutes), Uncertainty in adversarial learning (35 minutes), Concluding remarks (5 minutes).

Organiser Bio (Click to expand)Organiser Bio

Using intervals to capture and handle uncertainty

Organisers: Christian Wagner (<christian.wagner@nottingham.ac.uk>), Vladik Kreinovich (<vladik@utep.edu>), Shaily Kabir (<shaily.kabir@nottingham.ac.uk>), Zack Ellerby (<zack.ellerby@nottingham.ac.uk>).

Uncertainty is pervasive across data and data sources, from sensors in engineering applications to human perceptions (consumer preferences, patients’ assessment of pain, voters’ views in anticipation of elections), and the expertise of experts in areas as diverse as marketing to cyber security. Appropriate handling of such uncertainties depends upon three main stages: capture, modelling, and appropriate AI based handling of the uncertain information.
This tutorial will explain the nature of uncertainty at the level of individual data sources (such as people, sensors, etc.) highlighting the difference between between- (or inter-) and within- (intra-) source uncertainty. Focussing on the uncertainty arising from human-sourced data, the organisers will take participants through an end-to-end exercise in capturing and analysing interval-valued data, exploring the research question of ‘What are the attitudes to AI’ live with participants. Participants will specifically learn about intervals as a model to capture and model uncertainty, including the complexity and potential of associated techniques, from the basics (e.g. interval arithmetic), to current frontier topics underpinning the role of interval-valued data in AI (e.g. interval-valued regression). The tutorial is designed to be suitable for both students and academics/professionals from academia and industry without requiring prior expertise in the area.

This tutorial is designed to give researchers a practical introduction to the use of intervals for handling uncertainty in data. The tutorial will discuss relevant types and sources of uncertainty before proceeding to review and demonstrate practical approaches and tools that enable the capture, modelling and analysis of interval-valued data. This session will provide participants with an end-to-end overview and in-depth starting point for leveraging intervals within their own research.

Timing: The overall time of the tutorial will be three hours, with approximately 40 minutes per section, and a 20 minute break.

Pre-requisites: There are no pre-requisites for this tutorial, although a familiarity with fuzzy sets will be an advantage.

Organiser Bios (Click to expand)Organiser Bios