Curriculum learning for human compositional generalization

Dekker, Ronald; Otto, Fabian; Summerfield, Christopher

doi:10.1073/pnas.2205582119

Cited by 33 publications

(15 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As a consequence, participants could not predict the color sequence solely based on the location sequence in each trial. Finally, our study is also different from recent works on structure learning and generalization (Dekker et al, 2022;Garvert et al, 2017;Liu, Mattar, et al, 2021;Ren et al, 2022;Schapiro et al, 2013), as our task does not involve pre-exposure training or task-related rewards.…”

Section: Discussionmentioning

confidence: 95%

Shared structure facilitates working memory of multiple sequences

Huang

Luo

2023

Preprint

View full text Add to dashboard Cite

Daily experiences often involve the processing of multiple sequences, such as speech processing and spatial navigation, yet storing them challenges the limited capacity of working memory (WM). To achieve efficient memory storage, relational structures shared by sequences would be leveraged to reorganize and compress information. Here, participants memorized a sequence of items with different colors and spatial locations and later reproduced the full color and location sequences, one after another. Crucially, we manipulated the consistency between location and color sequence trajectories. First, sequences with consistent trajectories demonstrate improved memory performance and a trajectory correlation between the reproduced color and location sequences. Interestingly, color sequence undergoes spontaneous forward neural replay when recalling trajectory-consistent location sequence. These results reveal that shared common structure is spontaneously leveraged to integrate and facilitate WM of multiple sequences through neural replay and imply a role of common cognitive map in efficient information organization in WM.

show abstract

Section: Discussionmentioning

confidence: 95%

Shared structure facilitates working memory of multiple sequences

Huang

Luo

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…This provides a foundation for explaining how PFC representations should change with learning. Future experiments should extend this paradigm, to track changes in learning even more complex and naturalistic tasks 51 ; those that have a compositional structure 47,52 the influence of different learning curricula 53 ; and how these representations change within the same individual neurons as opposed to pseudo populations [54][55][56] .…”

Section: Discussionmentioning

confidence: 99%

Learning shapes neural geometry in the prefrontal cortex

Wójcik

Stroud

Wasmuht

et al. 2023

Preprint

View full text Add to dashboard Cite

The relationship between the geometry of neural representations and the task being performed is a central question in neuroscience. The primate prefrontal cortex (PFC) is a primary focus of inquiry in this regard, as under different conditions, PFC can encode information with geometries that either rely on past experience or are experience agnostic. One hypothesis is that PFC representations should evolve with learning from a format that supports exploration of all possible task rules to a format that minimises metabolic cost and supports generalisation. Here we test this idea by recording neural activity from PFC when learning a new rule (XOR rule) from scratch. We show that PFC representations progress from being high dimensional and randomly mixed to low dimensional and rule selective, consistent with predictions from metabolically constrained optimised neural networks. We also find that this low-dimensional representation facilitates generalisation of the XOR rule to a new stimulus set. These results show that previously conflicting accounts of PFC representations can be reconciled by considering the adaptation of these representations across learning in the service of metabolic efficiency and generalisation.

show abstract

“…Furthermore, our experiment was not designed to study whether and to what extent participants learn novel and potentially better planning strategies over time. Future studies might address learning dynamics, by comparing how people behave when presented with different sequences of planning problems – or “training curricula” (Dekker et al, 2022) – that afford or do not afford generalization across problems.…”

Section: Discussionmentioning

confidence: 99%

Adaptive planning depth in human problem solving

Eluchans

Lancia

Maselli

et al. 2023

Preprint

View full text Add to dashboard Cite

We humans are capable of solving challenging planning problems, but the range of adaptive strategies that we use to address them are not yet fully characterized. Here, we designed a series of problem-solving tasks that require planning at different depths. After systematically comparing the performance of participants and AI planners, we found that when facing manageable problems that require planning to a certain number of subgoals (from 1 to 6), participants make an adaptive use of their cognitive resources - namely, they tend to select an initial plan having the minimum required depth, rather than selecting the same depth for all problems. When facing more challenging problems that require planning up to 7 or 8 subgoals, participants tend to select a simpler (greedy) strategy wherein plans consider only the next 1 or 2 subgoals. Furthermore, we found a strong similarity between how different participants solve the same problems. These results support the view of problem solving as a bounded rational process, which adapts costly cognitive resources to task demands.

show abstract

Curriculum learning for human compositional generalization

Cited by 33 publications

References 39 publications

Shared structure facilitates working memory of multiple sequences

Shared structure facilitates working memory of multiple sequences

Learning shapes neural geometry in the prefrontal cortex

Adaptive planning depth in human problem solving

Contact Info

Product

Resources

About