|
|
|
Program > Accepted papers and talks
Invited Talks Gabriele Kern-Isberner Title: Cognitive Logics, and the Relevance of Nonmontonic Formal Logics for Human-Centred AI Abstract: Classical logics like propositional or predicate logic have been considered as the gold standard for rational human reasoning, and hence as a solid, desirable norm on which all human knowledge and decision making should be based, ideally. For instance, Boolean logic was set up as kind of an arithmetic framework that should help make rational reasoning computable in an objective way, similar to the arithmetics of numbers. Computer scientists adopted this view to (literally) implement objective knowledge and rational deduction, in particular for AI applications. Psychologists have used classical logics as norms to assess the rationality of human commonsense reasoning. However, both disciplines could not ignore the severe limitations of classical logics, e.g., computational complexity and undecidedness, failures of logic-based AI systems in practice, and lots of psychological paradoxes. Many of these problems are caused by the inability of classical logics to deal with uncertainty in an adequate way. Both disciplines have used probabilities as a way out of this dilemma, hoping that numbers and the Kolmogoroff axioms can do the job (somehow). However, psychologists have been observing also lots of paradoxes here (maybe even more). So then, are humans hopelessly irrational? Is human reasoning incompatible with formal, axiomatic logics? In the end, should computer-based knowledge and information processing be considered as superior in terms of objectivity and rationality? Cognitive logics aim at overcoming the limitations of classical logics and resolving the observed paradoxes by proposing logic-based approaches that can model human reasoning consistently and coherently in benchmark examples. The basic idea is to reverse the normative way of assessing human reasoning in terms of logics resp. probabilities, and to use typical human reasoning patterns as norms for assessing the cognitive quality of logics. Cognitive logics explore the broad field of logic-based approaches between the extreme points marked by classical logics and probability theory with the goal to find more suitable logics for AI applications, on the one hand, and to gain more insights into the rational structures of human reasoning, on the other.
Silja Renooij Title: Surfing the waves of explanation Abstract: The need for explaining black box machine learning models has revived the interest in explainability in AI more in general. One of the ideas underlying explainable AI is to use (new) models that are inherently explainable to replace or complement black box models in machine learning. Explainable or interpretable models exist and have existed for quite some time, and different aspects of these models and their outputs can be explained. In this talk I will focus on explanation of probabilistic graphical models, with an emphasis on Bayesian networks. From their very first introduction in the late 1980s, explanation of Bayesian networks has been a topic of interest: sometimes receiving a lot of attention, sometimes a seemingly forgotten topic, but now resurfacing again, riding on the waves of explainable AI.
Francesca Toni Title: Learning Argumentation Frameworks Abstract: Argumentation frameworks are well studied as concerns their support for various forms of reasoning. Amongst these, abstract argumentation and assumption-based argumentation framworks can be used to support various forms of defeasible, non-monotonic reasoning. In this talk I will focus on methods for learning these frameworks automatically from data. Specifically, I will overview two recent methods to obtain, respectively, abstract argumentation frameworks from past cases and assumption-based argumentation frameworks from examples of concepts. In both cases, the learnt frameworks can be naturally used to obtain argumentative explanations for predictions drawn from the data (past cases or examples) in the form of disputes, thus supporting the vision of data-centric explainable AI.
Tutorials Denis Bouyssou Title: How to use bibliometric indices (if you really must) Abstract: Higher education and research are often seen as affecting in a crucial way the economic performances of nations. Indeed, most countries devote a significant part of their resources to finance higher education and research institutions. Hence, we should expect that there is a growing tendency to evaluate and monitor their performances. Obviously, their very nature makes this task difficult and complex. We have recently witnessed a flourishing of evaluation agencies and a growing use of bibliometric indices of various kinds to evaluate individual scholars, departments, projects or universities. The aim of this tutorial is twofold. We will first outline the type of problems that may be encountered when evaluating research activities using standard bibliometric indices. We will then show how the classical tools provided by decision theory may be useful to analyze the theoretical properties of such indices. Our conclusion will be that some frequently used indices, such as the h-index, have rather undesirable properties.
Tanya Braun Title: A Glimpse into Statistical Relational AI: The Power of Indistinguishability Abstract: Statistical relational artificial intelligence, StaRAI for short, focuses on combining reasoning in uncertain environments with reasoning about individuals and relations in those environments. An important concept in StaRAI is indistinguishability, where groups of individuals behave indistinguishably in relation to each other in an environment. This indistinguishability manifests itself in symmetries in a propositional model and can be encoded compactly using logical constructs in relational models. Lifted inference then exploits indistinguishability for efficiency gains. This article showcases how to encode indistinguishability in models using logical constructs and highlights various ways of using indistinguishability during probabilistic inference.
Title: Decision under uncertainty Abstract: Decision under uncertainty is pervasive in artificial intelligence. The goal of this tutorial is to review some popular decision models and highlight their connections and the situations in which they can be applied. We will start with the expected utility model (EU) and show the properties it relies on. Then, we will present decision trees, that represent sequential decision problems, and show that the aforementioned properties allow for an efficient algorithm to solve them. Interpreting these trees differently will lead to another more efficient model called an influence diagram. The EU model is known to have severe limitations and, in some situations, more general decision models are needed. Based on a rephrasing of EU, we will present the more general rank dependent utility model (RDU). We will also show the issues of RDU w.r.t. sequential decision making. Another path to generalize EU is to substitute probabilities by other models, notably belief functions. We will show that the EU's properties presented at the beginning of the talk can also be applied to belif functions, hence resulting in the belief expected utility (BEU) model. We will conclude this talk, mentioning briefly other popular decision models.
Title: Data Lakes: a new Paradigm for Data Platforms and current Challenges Jean-Guy Mailly Title: On Incompleteness in Abstract Argumentation: Complexity and Expressiveness Abstract: One of the recent trends in research about abstract argumentation is the study of how incomplete knowledge can be integrated to argumentation frameworks (AFs). In this paper, we survey main results on Incomplete AFs (IAFs), following two directions: how hard is it to reason with IAFs? And what can be expressed with IAFs? We show that two generalizations of IAFs, namely Rich IAFs and Constrained IAFs, despite having a higher expressive power than IAFs, have the same complexity regarding classical reasoning tasks.
Michael Poss Title: An introduction to discrete robust optimization Abstract: Robust optimization (RO) has become a central framework to handle the uncertainty that arises in the parameters of optimization problems. The success of RO arises from its tractability since, unlike stochastic optimization, it does not suffer from the curse of dimensionality. This key property of RO leverages the structure of the uncertainty sets, which are described by small number of constraints constraints, often linear ones. In this tutorial, we will review these key aspects and cover two fundamental tractability results in discrete robust optimization. We will also illustrate these results on the knapsack problem and on the vehicle routing problem.
Jeremy Rohmer Title: Dealing with imperfect knowledge in natural hazard assessments: beyond classical probabilities and challenges Abstract: The distinction between two origins of uncertainty has become standard practice in risk analysis, namely random uncertainty (representing variability) and epistemic uncertainty (related to imperfect knowledge). While the former can be adequately represented using classical probabilities, there is no simple, single answer for the latter. New theories of uncertainty based on "imprecise probabilities" have been developed in recent years to go beyond the systematic use of a single probabilistic law. In this tutorial we analyse the advantages and disadvantages for the assessment of natural hazards (e.g. earthquakes, marine floods, landslides, etc.) with a comparison to the traditional probabilistic approach. We discuss the problems that have been solved and the interesting open questions and challenges that remain, in particular how to appropriately support decision making under uncertainty, how to provide guidance for future actions and how to deal with multiple forms of uncertainty along the assessment chain
Diedrich Wolter Title: Faithful Geometric Models for Integrating Learning and Reasoning Abstract: Knowledge graph embeddings are a direction in AI research that gained popularity due to its prospects of linking machine learning and concept-level logic reasoning. Linkage can be achieved by identifying geometric structures that a machine learning algorithm can construct and which then can serve as input to a geometrically grounded reasoning procedure. Already, knowledge graph embeddings have proven useful for link prediction, sometimes using geometric operations as simple as vector translations. However, the semantics of geometric models obtained by machine learning are fundamentally different from models in classical logics. This challenges the integration of learning and reasoning since the semantic gap needs to be bridged. In this presentation I review existing embedding techniques from the perspective of ontological and logical commitments they (implicitly) make, shedding some light on the semantic gap. I will then detail an approach based on the geometry of cones. Cones, if combined with a certain set of geometric operations, exhibit several interesting features. For example, cones allow us to build faithful geometric models for semi-expressive concept languages that retain uncertainty present in training data. In the light of these findings I motivate further investigations of geometric structures for learning, representation, and reasoning.
Accepted papers :
|
| Online user: 2 | Privacy |
|
