Krishn Bera

How fast we decide reveals how we learn. New in @plos.org #PLoSCompBio : modeling response times alongside choices disentangles working memory, reinforcement learning and cognitive control, and reveals differences in schizophrenia that choices alone missed. 🧵 t.co/BTOgmpd4YH

View full thread

For computational psychiatry: joint modeling of choices and response times enables accurate, model-based and mechanism-oriented computational phenotyping. The approach is broadly extensible, letting researchers compare theories of learning and decision dynamics within one model.

Joint work with Alexander Fengler and Michael Frank, and our CNTRaCS colleagues Megan Boudewyn, Cameron Carter, Molly Erickson, Jim Gold, Steve Luck, Dan Ragland, Andy Yonelinas, Angus MacDonald and Deanna Barch. @lnccbrown.bsky.social #CogSci #CompCogNeuro #Psychiatry

View full thread

In schizophrenia, the model confirmed previously known working memory deficits and revealed two novel differences hidden in earlier choice-based analyses: slower RL/trial-and-error learning and lesser extent of this adaptive slowing strategy.

The key takeaway: response times are informative about how we learn. Most learning models discard them, but modeling decision dynamics enabled better characterization of RL, WM and cognitive control contributions to learning behaviors.

View full thread

Models fit to choices alone overestimate how quickly people learn by trial and error. We validate this on a later, out-of-sample test phase the model never saw, where memory cannot help. PADDL predicted people's choices there; the choice-only model was overconfident.

Response times also revealed an adaptive strategy: when people know a harder task block with greater WM load is coming, they slow down from the very start to be more accurate. A signature of proactive cognitive control, invisible in choices alone.

When we learn from feedback, two systems work in parallel: working memory, fast but limited to a few items, and reinforcement learning, slow but durable. They usually point to the same answer, so modeling choices alone blur their contributions.