- Title Information
- Title
- Computational and Behavioral Mechanisms of In-Context Learning
- Type of Resource (primo)
- dissertations
- Name:
Personal
- Name Part
- Rashed Ahmed, Abdullah P
- Role
- Role Term:
Text
- creator
- Name:
Personal
- Name Part
- Sheinberg, David
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Serre, Thomas
- Role
- Role Term:
Text
- Advisor
- Name:
Personal
- Name Part
- Nassar, Matthew
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Frank, Michael
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Gold, Joshua
- Role
- Role Term:
Text
- Reader
- Name:
Corporate
- Name Part
- Brown University. Department of Neuroscience
- Role
- Role Term:
Text
- sponsor
- Origin Information
- Copyright Date
- 2026
- Physical Description
- Extent
- xvi, 136 p.
- digitalOrigin
- born digital
- Note:
thesis
- Thesis (Ph. D.)--Brown University, 2026
- Genre (aat)
- theses
- Abstract
- Sensory features of the environment change continuously, often in context-dependent ways that alter their underlying statistical structure. Recognizing and responding appropriately to such changes requires organisms to identify and track features of the local context, using this information to interpret ongoing sensory experience and update their predictions accordingly. To investigate this, we introduce a novel predictive inference paradigm, the Bouncing Ball task, in which observers predict the partially observable color of a moving ball. Critically, the probability of a color change depends on two latent variables that participants must estimate on a trial-by-trial basis. Using the ideal Bayesian observer, we define the optimal behavioral signatures of in-context learning and demonstrate that both humans and recurrent neural networks trained on the task, exhibit these signatures. However, we observe asymmetries in learning outcomes for humans and models, suggesting differences in capacity for evidence accumulation of each latent variable. To identify the computational mechanisms learned by the recurrent models, we probed their learned representations using elastic net regularization, temporal dynamics analysis, and targeted state interventions to identify the minimal neural sub-circuit that drives adaptive, in-context learning. These analyses revealed that just a single LSTM unit encoded each task parameter through functionally independent leaky integrator circuits. Our findings reveal fundamental computational components necessary for in-context learning and provide a bridge between normative Bayesian theories of human flexibility, interpretable dynamics in recurrent neural networks, and human neurophysiology.
- Subject
- Topic
- behavioral neuroscience
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00872004")
- Topic
- Computational neuroscience
- Language
- Language Term (ISO639-2B)
- English
- Record Information
- Record Content Source (marcorg)
- RPB
- Record Creation Date
(encoding="iso8601")
- 20260516