An Invitation to Analogies in (A)I
With machine learning models taking over more tasks from biological intelligences, there is a renewed interest in characterising fundamental properties of intelligence itself. In this workshop, we will investigate one such aspect: analogical thinking and reasoning.
Even before artificially intelligent systems were integrated with machine learning techniques, analogies and abstractions were recognised as important aspects of intelligence. Famously, Douglas Hofstadter argued that analogical reasoning forms the core of human cognition. While this intuition is still common, identifying and interpreting reasoning in opaque machine learning systems hinders a deeper understanding. However, as modern AI systems increasingly engage with complex, open-ended tasks, their ability to engage in analogical reasoning becomes crucial for their performance and interpretability, which inspired this workshop’s theme.
We will hear different perspectives on analogies, intelligence, explore the tension between symbolic and connectionist views, and more. ​
Confirmed invited speakers
Melanie Mitchell
(Santa Fe Institute)
Jules Hedges
(Institute for Categorical Cybernetics & University of Strathclyde)
Martha Lewis
(University of Amsterdam)
Han van der Maas
(University of Amsterdam & Santa Fe Institute)
Schedule and Registration
This workshop is a full-day event, starting at 9:30 and ending around 16:30.
​
Location:
The location of the workshop is Mediamatic, Dijksgracht 6, 1019 BS Amsterdam.
​
Registration:
Please register through this form.
​
Program:
9:30: Arrival, with coffee and tea
​
10:00-10:15: Welcome and introduction
10:15-11:15: Melanie Mitchell
11:15-12:15: Han van der Maas
12:15 - 13:15: Lunch break
13:15 - 14:15 Jules Hedges
14:15 - 14:30: Short break
14:30 - 15:30: Martha Lewis
​
15:30 - 16:30: Drinks and goodbye
​
Talks:
Melanie Mitchell — Analogy and Abstraction In Humans and Machines
Conceptual abstraction and analogy-making are key abilities underlying humans' abilities to learn, reason, and robustly adapt their knowledge to new domains. In this talk I will describe attempts to imbue such abilities in AI systems, as well as methods for evaluating the robustness of such abilities in both machines and humans.
Han van der Maas — Multiple crosspoints in the study of natural and artificial intelligence
I start with a historical overview of a century of research on human intelligence. To understand this trajectory—including its scientific advances and ethical controversies—I draw on Cronbach’s influential distinction between two psychological traditions: the study of individual differences and the study of mental processes. I then trace the development of artificial intelligence and its entanglement with theories of human cognition, using chess as a guiding example. I conclude with a critical analysis of how intelligent today’s AI systems truly are.​​
​
Jules Hedges — Bidirectional abstraction in symbolic AI and machine learning
I will argue that the tension between symbolic (aka. good old-fashioned) and quantitative (aka. connectionist or machine learning) methods in AI is greatly exaggerated and should be seen as different parts of an abstraction hierarchy. I will propose a mathematical gadget called "lenses" as a common theory of the bidirectional processes of abstraction and specification that appears in both branches of AI, giving a possible starting point for building bridges between them.
​
Martha Lewis — Behavioural vs. Representational Systematicity in AI Systems
The ability to systematically and productively recombine attributes, concepts, and rules is a key characteristic of human intelligence. Systematic behaviour is highly sought after in AI systems, however, assessing systematicity only at the behavioural level can lead to brittleness under changes of context. Rather, systematicity at the representational level is needed. This talk will discuss different ways of building or discovering systematic representations in AI systems, and how these different methods could be related.
