Michael J. Frank, PhD
Edgar L. Marston Professor of Cognitive & Psychological Sciences
Director, Carney Center for Computational Brain Science
Brown University
Host(s): Kaushik Lakshminarasimhan (Postdoc)
Dynamic interplay between In-context and in-weight learning in humans and neural networks
Human learning embodies a striking duality: sometimes, we appear capable of following logical, compositional rules and benefit from structured or blocked curricula, while other times, we rely on an incremental approach or trial-and-error, learning better from curricula that are unstructured or randomly interleaved. Influential psychological theories explain this seemingly disparate behavioral evidence by positing two qualitatively different learning systems -- one for rapid, rule-based inferences and another for slow, incremental adaptation. It remains unclear how to reconcile such theories with neural networks, which learn via incremental weight updates and are thus a natural model for the latter type of learning, but are not obviously compatible with the former. However, recent evidence suggests that both metalearning neural networks and large language models are capable of "in-context learning" (ICL) -- the ability to flexibly grasp the structure of a new task from a few examples given at inference time. Here, we show that networks capable of ICL can reproduce human-like learning and compositional behavior on rule-governed tasks, while at the same time replicating human behavioral phenomena in tasks lacking rule-like structure via their usual in-weight learning (IWL). Our work shows how emergent ICL can equip neural networks with fundamentally different learning properties than those traditionally attributed to them, and that these can coexist with the properties of their native IWL, thus offering a novel perspective on dual-process theories and human cognitive flexibility, with links to working memory and reinforcement learning.
Relevant Publications:
Chunking as a Rational Strategy for Lossy Data Compression in Visual Working Memory
Venue Information:
Speaker Location: Jerome L. Greene Science Center, Kavli Auditorium, 9th Floor Lecture Hall
The Columbia Neuroscience Seminars have been organized to help build community and collaboration among researchers interested in this broad field across campus. The in-person activities, including the talks, provide meaningful interactions for the speakers, many of whom have traveled a long way to visit Columbia. However, if you are a Columbia researcher on another campus and are unable to attend the talk at the Jerome L. Greene Science Center, please email [email protected] at least 48 hours in advance to request an individual, one-time Zoom link (livestream only, no Q&A).
If you have a short- or long-term accommodation request (medical issue, travel, other concerns, etc.), or any other questions, please also reach out to [email protected].
Tuesdays@10 is a signature Zuckerman Institute initiative that aims to expose researchers at all levels to high-quality science and stimulate scientific discourse. The speakers featured in this series represent various fields and techniques in neuroscience, and include invited guests of the Columbia Neuroscience Seminars, the Zuckerman Institute's Local Circuits Affiliates Program, and other special seminar series through a combined, collaborative effort of one or more of the following: Columbia's Zuckerman Institute, the Center for Precision Psychiatry, the Department of Neuroscience, the Doctoral Program in Neurobiology and Behavior and the Columbia Translational Neuroscience Initiative, and with support from the Kavli Institute for Brain Science.
More information and a full schedule can be found here.