Organizers:
Taylor Webb (University of California, Los Angeles)
Claire Stevenson (University of Amsterdam)
Speakers
Taylor Webb (University of California, Los Angeles)
Ellie Pavlick (Brown University)
Eunice Yiu (University of California, Berkeley)
Martha Lewis (University of Bristol)
Moderator
Claire Stevenson (University of Amsterdam)
Taylor Webb (University of California, Los Angeles)
Bio: Taylor Webb is a postdoctoral researcher in the Department of Psychology at UCLA (The University of California, Los Angeles). His research is focused on understanding the computational mechanisms that support higher-order cognition (e.g., analogical reasoning, metacognition) in both the human brain and artificial systems.
Talk Title: Implications of Emergent Analogical Reasoning for Cognitive Science
Abstract: Has analogical reasoning emerged in large language models (LLMs), and what, if anything, might this teach us about human analogical reasoning? In this talk, I will review recent evidence that LLMs are capable of solving a wide range of analogy tasks, including some adversarial tasks explicitly designed to be dissimilar to their training data, and will argue that these findings are best explained by the emergence of core mechanisms for analogical reasoning (e.g., analogical mapping, schema induction). I will also caution against overgeneralization of these findings, highlighting ways in which analogy interacts with other processes in which LLMs display notable weaknesses, including physical reasoning and multi-object vision. Finally, I will argue that there is still much important work to be done in cognitive science – LLMs have not ‘solved’ analogy. Indeed, the quest to understand how these systems reason, and how this might relate to human reasoning, has arguably just begun, and the cognitive science of analogy has an important role to play in this effort.
Eunice Yiu (University of California, Berkeley)
Bio: Eunice Yiu is a fifth-year PhD student at UC Berkeley. She is advised by Dr. Alison Gopnik in the Department of Psychology. She studies object perception, exploration
and innovation in children and AI, as well as the causal learning and analogical reasoning underlying these cognitive processes.
Talk Title: Are Visual Analogies Really Solved? Decoding Reasoning Across Children, Adults, and Foundational Models
Abstract: Humans are robust and flexible learners, adept at adapting to nonstationary environments and novel situations through analogical reasoning. Current large
pretrained models excel in specific visual tasks, but standard benchmarks fail to evaluate their basic visual analogical reasoning capacity. We developed a benchmark
featuring 2,900 common visual transformations of everyday objects, inspired by human developmental psychology. We compare the performance of large multimodal models (LMMs) compared to that of human adults and children on our benchmark. Our evaluation includes three stages: detecting the domain of visuospatial transformation, generalizing the transformation rule, and predicting or extrapolating the rule to novel scenarios. Findings show that GPT-4V, LLaVa-1.5, and MANTIS excel in early reasoning stages but struggle with precise transformations and visual extrapolation to new objects outside the training data. By contrast, humans (even young children) outperform models in detecting, generalizing, and extrapolating visual transformation rules, highlighting the limitations of models trained primarily on 2D images and text.
Martha Lewis (University of Bristol)
Bio: Martha is a Lecturer in the School of Engineering Mathematics and Technology at the University of Bristol, UK. Prior to Bristol, she held a Veni fellowship at the University of Amsterdam, was a postdoc in the Department of Computer Science at the University of Oxford, and currently is visiting the Santa Fe Institute working on approaches to modelling analogy. Her research interests are in compositional approaches to modelling language and concepts, and in how these can be integrated with neural or distributed systems.
Talk Title: Evaluating the Generality of Analogical Reasoning in Large Language Models
Abstract: Large language models (LLMs) have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, it has been debated whether they are actually performing humanlike abstract reasoning or instead employing less general processes that rely on similarity to what has been seen in their training data. I will discuss the robustness of analogy-making abilities previously claimed for LLMs, using “counterfactual'' variants of analogy problems---versions that test the same abstract reasoning abilities but that are likely dissimilar from any pre-training data. I will compare performance of humans and three GPT models on both the original and counterfactual problems, and show that, while the performance of humans remains high for all the problems, the GPT models' performance declines sharply on the counterfactual set. This provides evidence that, despite previously reported successes of LLMs on analogical reasoning, these models lack the robustness and generality of human analogy-making.
Ellie Pavlick (Brown University)
Bio: Ellie Pavlick is an Associate Prof. of Computer Science and Linguistics at Brown University and a Research Scientist at Google. She received her PhD from University of Pennsylvania in 2017, where her focus was on paraphrasing and lexical semantics. Her work focuses on computational models of language (currently, primarily LLMs) and its connections to the study of language and cognition more broadly. Ellie leads the language understanding and representation (LUNAR) lab, which collaborates with Brown’s Robotics and Visual Computing labs and with the Department of Cognitive, Linguistic, and Psychological Sciences.
Talk Title: Reasoning In-Context and In-Weights in LLMs
Abstract: Reasoning by analogy requires combining flexible in-context inferences with rich knowledge of past domains and experiences stored in model weights. The mechanisms that govern in-context vs. in weights learning and use within LLMs are as yet poorly understood. I will survey a number of recent studies in my lab, focusing first on examples of LLMs reasoning successfully about analogies, but failing to mirror human learning and error patterns. Then, I will describe several results on LLM mechanisms for in context and in-weights learning and reasoning. I will end with speculation on how these lines of work can combine to advance our understanding of analogical reasoning in humans and machines.
Claire Stevenson (University of Amsterdam)
Bio: Claire Stevenson is an assistant professor at the University of Amsterdam's Psychological Methods department, with a background in Developmental Psychology and Computer Science. Her research focuses on human versus AI intelligence and creativity, with a special interest in the development of analogical reasoning in children vs AI models.
Organizer
Nick Ichien (University of Pennsylvania)
Speakers
Rachel Flood Heaton (University of Illinois Urbana-Champagne) - Modeling the perception of relational structure in vision
Cas Coopmans (Donders Centre for Cognitive Neuroimaging) - Two challenges for cognitive computational models of language
Abhishek Dedhe (Carnegie Melon University) - Cognitive algorithms underlying hierarchical pattern generation across age and species
Bonan Zhao (Princeton University) - Constructing deep concepts through shallow search
Moderator
Nick Ichien (University of Pennsylvania)
Rachel Flood Heaton (University of Illinois Urbana-Champagne) - Vision
Bio: Rachel Heaton is an Intelligence Community Postdoctoral Research Fellow hosted by the New Frontiers Initiative at the University of Illinois Urbana-Champaign. Her research program integrates the fields of human-centered design, cognitive psychology, and computational modeling. The goal of her research is to better understand how people mentally represent and reason about objects they see in the world and how perception and cognition impact human responses to designed artifacts. Her scientific work includes computational models of object recognition, visual attention, reasoning via visual analogy, and the perception of affordances, as well as the study of general issues in neural computation and artificial intelligence. She received her PhD from the University of Illinois Urbana-Champaign in the psychology of visual attention & perception. She holds a Master of Fine Arts in industrial design and a Bachelor of Science in electrical engineering from the University of Illinois Urbana-Champaign. Prior to her research in design, psychology, and neural representation, she worked in the semiconductor industry designing digital logic for CPU and network processor architectures.
Talk Title: Modeling the perception of relational structure in vision
Abstract: Visual reasoning plays a prominent role in human mental life and depends on our ability to perceive relational structures. Modeling visual structure in neural networks is a challenge. Some models do a good job of predicting data but are not image computable; others are image computable but cannot properly represent relations. I will present work from an ongoing effort to solve both problems by combining the JIM model of object recognition (Hummel & Biederman, 1992) and the LISA model of analogical reasoning (Hummel & Holyoak, 1997; 2003). The two models use a common representational scheme, making it possible to use JIM as a visual input to LISA. We have built a system to serve as an interface between JIM and LISA, which already accounts for several phenomena in the perception of structure. Toward the aim of making the system image computable and neurally realistic, we are augmenting the principles in JIM with known properties of human striate cortex. In addition to aspects of perceptual grouping, this model accounts for bottom-up visual salience with very few free parameters. Our aim is to build a more complete model of ventral vision that can serve as a visual interface for an analogical reasoning engine.
Cas Coopmans (Donders Centre for Cognitive Neuroimaging) - Language
Bio: Cas Coopmans is a postdoctoral researcher at the Donders Centre for Cognitive Neuroimaging in Nijmegen, The Netherlands. In his work he uses behavioral, electrophysiological (EEG, MEG), and computational methods to study the nature of linguistic representations in the human brain. Previously, during his PhD at the Max Planck Institute for Psycholinguistics, he studied the role hierarchical structure in language use.
Talk title: Two challenges for cognitive computational models of language
Abstract: Over the past several years, it has been shown that artificial neural network (ANN) models can learn to generate many complex syntactic constructions, seemingly without relying on the type of symbolic structured knowledge conventionally used by language scientists. However, the evaluation criteria for measuring the linguistic capacities of ANNs are often different from those used for inferring human linguistic competence, lending some to suggest that the linguistic representations of ANNs are not human-like. In the context of this misalignment, I will present two challenges for cognitively-oriented computational models of language. My arguments are embedded in a discussion of classic arguments for hierarchical structure in syntax, and empirical evidence from behavioral and computational experiments that compare the linguistic performance of humans and ANNs on the same linguistic tasks.
Abhishek Dedhe (Carnegie Mellon University) - Reasoning
Bio: Abhishek Dedhe is a final year PhD student in Dr. Jessica Cantlon's lab at Carnegie Mellon. He is pursuing a PhD in Cognitive Neuroscience and the Quantitative Certificate for Psychology and Neuroscience. He studies the origins of hierarchical reasoning, a cognitive ability seen in abstract thought and complex-problem solving domains such as recursion in language, mathematics, music, and motor behaviour. He collects and analyzes behavioral data from healthy human adults and 3-6 year old children as well as conducting new analyses with previously published data from monkeys and crows; these analyses are performed using a Bayesian mixture model.
Talk title: Cognitive algorithms underlying hierarchical pattern generation across age and species
Abstract: Humans are powerful generalizers capable of inferring, predicting, and generating patterns across many domains. Hierarchical patterns are composed of lower-level units combined to form higher-level ones. In particular, the capacity to process recursive hierarchical patterns — consisting of structures embedded within other structures of the same kind — is speculated to be the key feature that distinguishes human cognition from animal cognition. Some theories propose that recursion and symbolism are innate traits in humans (Deheane et al., 2015; Fitch, 2014). Another possibility is that the distinction between humans and non-humans lies in domain-general differences in information-processing capacities, such as working memory, required for complex rules, hierarchical patterns, and abstract generalizations (Cantlon & Piantadosi, 2024; Greenfield, 1991; Ferrigno et al., 2020; Halford, 1993). Here, I present results from a hierarchical patterning task (Ferrigno et al., 2020) where participants learn and generalize recursive center-embedded sequences. I used a Bayesian mixture model (Dedhe et al., 2023) to calculate p(cognitive algorithm | behavior) — the inferred probability that a given cognitive algorithm (such as associative chaining, linear ordering, and hierarchical embedding) was involved in generating complex behavior. I implemented cognitive algorithms as domain-general pattern-processing models capable of finding precise descriptions of the observed data. These algorithms were implemented under a simple Language of Thought framework and their computational complexity was measured using minimum description length, a common information-theoretic metric. In line with previous findings (Ferrigno et al., 2020; Liao et al., 2022), I test whether non-humans use hierarchical cognitive algorithms to generate hierarchical patterns and whether there are continuous, quantifiable advantages for humans (including children) relative to non-human performance. Taken together with previous work, these results refute a strong version of the “Dendrophilia” hypothesis, suggesting that humans are not unique in the ability to learn and generate hierarchical recursive patterns. Finally, I examine the relation between the computational complexity of the cognitive algorithms an individual can deploy and the global information processing capacity of the individual.
Bonan Zhao (Princeton University) - Conceptual learning
Bio: Bonan Zhao is a postdoctoral researcher at Princeton University, studying the computational principles driving people’s conceptual discoveries. She obtained her PhD at the University of Edinburgh, and Master of Logic at the University of Amsterdam. Before that, she was a philosophy major in Tsinghua University in Beijing, China, and spent a few years working as a data scientist in Amsterdam.
Talk title: Constructing deep concepts through shallow search
Abstract: The world is full of unseen, unknown and unexplored realms, yet people can effectively navigate through these novel situations, in part by applying causal relationships learned from the known situations to the unseen ones. However, the stark contrast between the complexity of our conceptual systems and how bounded our cognition is makes one wonder: How are people able to create and grasp such complex concepts that seem so far beyond their reach? Drawing from symbolic generative models and Bayesian nonparametric inference, I will argue that to achieve such ability, it requires two key components: (1) a structured representational substrate, and (2) an effective cache-and-reuse mechanism. The first component enables data-efficient learning and robust generalization, and the second component makes a resource-rational agent to go beyond its limitations. This model predicts systematically different learned concepts when the same information is processed in different orders, as strongly supported by a series of behavioral experiments. I will conclude by contemplating how bootstrap learning may serve as a computational account for why human learning is by nature modular, incremental, and fundamentally diverse.
Nick Ichien (University of Pennsylvania) - Moderator
Bio: Nick is a postdoctoral researcher at the University of Pennsylvania advised by Sudeep Bhatia, and he received his PhD in cognitive psychology from UCLA. His work aims to clarify the mental processes constituting human higher cognition and the representations over which they operate. Specifically, he is interested in human processing of analogy, causality, metaphor, similarity, sameness and difference, as well as memory search, recall, and recognition.
Organizer
Lindsey Richland (University of California, Irvine)
Speakers
Dedre Gentner (Northwestern University)
Jean-Pierre Thibaut (University of Bourgogne)
Micah Goldwater (University of Sydney)
Lindsey Richland (University of California, Irvine)
Moderator
Jean-Pierre Thibaut (University of Bourgogne)
Dedre Gentner (Northwestern University)
Bio: Dedre Gentner is the Alice Gabrielle Twight Professor of Psychology and Education at Northwestern University, a Fellow of the Cognitive Science Society, a Fellow of the American Academy of Arts and Sciences, and a member of National Academy of Sciences. She has received numerous awards for her work including the Alexander von Humboldt Research Award in 2013, the APA Distinguished Scientific Contributions in 2016, and the David E. Rumelhart Prize for Contributions to the Theoretical Foundations of Human Cognition in 2016. Professor Gentner is best known for her work in analogical cognition, and in particular, for developing the influential Structure-mapping Theory of analogical processing. Within this theoretical framework, she has researched a wide range of psychological phenomena, including concept and language learning, analogical development, metaphor processing, scientific discovery, and the relationship between language and higher-order cognition. In collaboration with Ken Forbus, she also co-developed the Structure-mapping Engine, a computational instantiation of Structure-mapping Theory.
Talk Title: Analogical Processes in Learning and Education
Abstract: Basic research on analogical processes has repeatedly shown that analogical comparison fosters learning and discovery. Analogical processing can highlight common relational systems, promote new inferences, and reveal potentially important differences between situations. This suggests that analogy could be powerful in education. In this talk, I’ll describe work on spatial analogy, causal analogy and mathematical analogy to show how basic research in analogical mapping naturally translates into educationally relevant work.
Jean-Pierre Thibaut (University of Bourgogne)
Bio: Jean-Pierre Thibaut is Professor of Developmental Psychology at the University Burgundy, Franche-Comté, France. His research focuses on the interwoven processes of concept and word learning, generalisation and analogical reasoning development. He has a specific interest in the role that executive functions play in these developmental processes and the use of eye-tracking methods to elucidate how executive functions interact with reasoning processes.
Talk Title: Multiplying sausages and tulips or tulips and daffodils are different math. Mapping the world on algorithms
Abstract: Studies on mathematical competence relate math with analogies. For example, comparing two fractions (e.g., 3/2 is equivalent or not to 6/4) is establishing that they are built around the same ratio (De Wolf et al., 2016; Starr et al. 2023). Here we illustrate that doing math is to connect (successfully or not) the world of mathematics (algorithms) and the world of objects on which the algorithms operate (tulips, sausages). We will illustrate this with the case of distributive problems to be solved by either factorization or development. We show that the algorithm choice depends on the nature of objects used in the problem, though this is mathematically irrelevant. This effect is present in both adults and children. We suggest that doing math goes beyond math and involves the nature of relations between the objects.
Micah Goldwater (University of Sydney)
Bio: Dr. Micah B. Goldwater is a Senior Lecturer in The School of Psychology at The University of Sydney. He researches how people think and learn in the lab, the classroom, and unfortunately on the internet.
Talk Title: The need to train elaborative learning strategies: evidence from category, and other kinds of, learning experiments
Abstract: There have historically been assumptions that formal learning models capture universal cognitive processes. However, for the past decade there has been growing evidence that different models may be capturing distinct intentional learning strategies. In this walk, I will present evidence that 1. there are consistent individual differences in learning strategy across a wide range of laboratory learning tasks and in undergraduate coursework 2. the strategy difference may be the volitional use of elaborative learning processes such as analogical comparison 3. the strategy difference is not associated with differences in cognitive capacity such as fluid intelligence or working memory. After laying out the evidence, I will discuss possibilities for how we could potentially train students to choose these strategies when self-regulating their own learning. Because experimental work shows we can prompt the use of analogical comparison and other elaborative learning strategies, this suggests it should be possible to train people to do so, yet currently evidence for this is lacking.
Lindsey Richland & Hongyang Zhao (University of California, Irvine)
Bio: Dr. Lindsey Richland is a psychologist whose research centers on analogical thinking and learning, with a particular focus on learning in the context of mathematics. She is a Professor of Education at the University of California, Irvine, previously at the University of Chicago. Her work has received regular federal funding from the U.S. Institute of Education Sciences, National Science Foundation, National Institute of Health, and the Office of Naval Research, and she publishes frequently in both psychological and educational fields. Dr. Richland uses a range of methodologies to understand how children develop the capacity and skills for analogical thinking, as well as to discover ways to optimize student learning from analogy in classroom settings.
Dr. Hongyang Zhao is a postdoctoral researcher at the Science of Learning Lab, University of California, Irvine. She recently graduated from the Department of Human Development and Quantitative Methodology at University of Maryland with a PhD in Educational Psychology and a master-level certificate in Measurement, Statistics, and Evaluation. Her research focuses on higher order thinking including relational thinking/reasoning, measurement issues (measure development, revision, and validation), and meta-analysis.
Talk Title: Learning to Notice to Learn by Analogy
Abstract: Learning by analogy is a powerful tool for drawing connections, making generalizations across contexts, and building structured representations of domains such as mathematics, yet learning from analogies requires noticing relational correspondences, and recognizing opportunities for drawing inferences based on these alignments. I describe research within the mathematics domain that highlights variability in children's noticing of relational correspondences and mathematical quantitative relations, suggesting this variability can be driven by individual differences in children, external factors such as feelings of pressure, or cues within the instructional materials themselves - signaling the need for educators to be intentional about their visual and verbal cues to help all students notice the need to learn from analogy.