@article{bibcite_202871,
  keywords = {motor sequence learning, context-dependent memory, similarity, neural networks},
  author = {Mohan Gupta and Jordan A. Taylor},
  title = {Lumping or Splitting: How Context Shapes Motor Sequence Representations},
  abstract = {<p>Motor skills are often assumed to rely on a shared representation that generalizes across superficial changes in framing. Yet identical actions are practiced under distinct contexts that may gate what is retrieved and expressed. We tested whether performance{\textendash}inconsequential context pushes learning toward generalization ({\textquotedblleft}lumping{\textquotedblright}) or context{\textendash}specific codes ({\textquotedblleft}splitting{\textquotedblright}). Simple recurrent networks (SRNs) learned nextstep prediction for a fixed eight{\textendash}item sequence in no{\textendash}context, single{\textendash}context, or alternating{\textendash}context conditions, with weight scale manipulated to bias the learned solution while holding the SRN architecture fixed. Humans trained on the same sequence under matched contexts. All groups improved with practice, but alternating contexts produced a sustained acquisition cost and a marked transition cost when switching between contexts. Under identical training, Low{\textendash}Scale SRNs showed crosscontext generalization, whereas High{\textendash}Scale SRNs reproduced the context{\textendash}dependent costs observed in human behavior. White{\textendash}box analyses revealed separable context{\textendash}specific representations and greater weighting of context inputs in High{\textendash}Scale SRNs, consistent with representational splitting.\&nbsp;</p>},
  year = {2026},
  journal = {Proceedings of the 48th Annual Conference of the Cognitive Science Society},
  chapter = {175},
  pages = {175-182},
  url = {https://escholarship.org/uc/item/0q05h85q},
}