Trial-by-trial learning of successor representations in human behavior DOI Creative Commons
Ari E. Kahn, Danielle S. Bassett, Nathaniel D. Daw

и другие.

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Ноя. 7, 2024

Abstract Decisions in humans and other organisms depend, part, on learning using models that capture the statistical structure of world, including long-run expected outcomes our actions. One prominent approach to forecasting such is successor representation (SR), which predicts future states aggregated over multiple timesteps. Although much behavioral neural evidence suggests people animals use a representation, it remains unknown how they acquire it. It has frequently been assumed be learned by temporal difference bootstrapping (SR-TD(0)), but this assumption largely not empirically tested or compared alternatives eligibility traces (SR-TD( λ > 0)). Here we address gap leveraging trial-by-trial reaction times graph sequence tasks, are favorable for studying dynamics because long horizons these studies differentiate transient update different rules. We examined behavior SR-TD( ) probabilistic task alongside number alternatives, found was best explained hybrid model via an additional zeroth-order predictive model. The relatively large estimate indicates predominant role trace mechanisms bootstrap-based chaining typically assumed. Our results provide insight into learn representations, demonstrate simultaneously SR lower-order predictions.

Язык: Английский

Structural knowledge: from brain to artificial intelligence DOI Creative Commons
Yingchao Yu, Yuping Yan, Yaochu Jin

и другие.

Artificial Intelligence Review, Год журнала: 2025, Номер 58(9)

Опубликована: Июнь 4, 2025

Язык: Английский

Процитировано

0
Trial-by-trial learning of successor representations in human behavior DOI Creative Commons
Ari E. Kahn, Danielle S. Bassett, Nathaniel D. Daw

и другие.

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Ноя. 7, 2024

Abstract Decisions in humans and other organisms depend, part, on learning using models that capture the statistical structure of world, including long-run expected outcomes our actions. One prominent approach to forecasting such is successor representation (SR), which predicts future states aggregated over multiple timesteps. Although much behavioral neural evidence suggests people animals use a representation, it remains unknown how they acquire it. It has frequently been assumed be learned by temporal difference bootstrapping (SR-TD(0)), but this assumption largely not empirically tested or compared alternatives eligibility traces (SR-TD( λ > 0)). Here we address gap leveraging trial-by-trial reaction times graph sequence tasks, are favorable for studying dynamics because long horizons these studies differentiate transient update different rules. We examined behavior SR-TD( ) probabilistic task alongside number alternatives, found was best explained hybrid model via an additional zeroth-order predictive model. The relatively large estimate indicates predominant role trace mechanisms bootstrap-based chaining typically assumed. Our results provide insight into learn representations, demonstrate simultaneously SR lower-order predictions.

Язык: Английский

Процитировано

0