Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Human Single Neurons and Predictive Models Remap Differently in Reward and Transition Relearning

Weijia Zhang¹, Sandra Maesta-Pereira¹, Tom Donoghue², Gelana Tostaeva³, Oscar Araiza Carranza⁴, Bradley Lega⁴, Ignacio Saez³, Jesse P Geerts⁵, Kim Stachenfeld¹, Joshua Jacobs⁶; ¹Columbia University, ²University of Manchester, ³Icahn School of Medicine at Mount Sinai, ⁴UT Southwestern Medical Center, ⁵Imperial College London, ⁶University of Chicago

Presenter: Weijia Zhang

Humans exhibit remarkable flexibility in goal-directed navigation, readily adapting to changes in spatial environments and associated rewards. The hippocampus is hypothesized to support this flexibility by encoding predictive relationships between states and rewards. We evaluated human behavior and single-neuron representational dynamics during a spatial sequence learning task that induces relearning through either transition revaluation (TR; changes in spatial structure) or reward revaluation (RR; changes in terminal rewards). To model these capabilities, we used neural sequence models, which learn predictive representations by associating non-consecutive spatial experiences as a framework for hippocampal-like predictive associative learning. At the behavioral level, prior work shows greater adaptation in RR than TR, suggesting that humans differentially update representations across conditions: reward updates are local, whereas transition updates rely on non-local temporal statistics that are slower to update. Our patient cohort (n = 16) exhibited a similar asymmetry. To model the effects of non-local spatial associations, we evaluated transformers with varying attention masks alongside RNNs. Across models, greater reliance on non-local predictive spatial associations led to a larger RR–TR performance gap. At the neural level, population activity in human neurons dissociated reward and transition relearning, as well as behavioral outcomes (successful vs. unsuccessful reevaluation), reflecting orthogonal representational updates. Both human neurons and sequence models exhibited greater spatial remapping during TR than RR. Together, these findings reveal a stability–flexibility trade-off in intelligent systems: transition representations encode non-local spatial relationships across states and are therefore slower and more costly to update than reward representations.

Topic Area: Memory, Learning & Knowledge Structures