Fig. 6: Schematic of computational processes underlying rule inference.
From: Abstract rule learning promotes cognitive flexibility in complex environments across species
Value-based strategy selection and top-down attention affect decision making and learning in a trial. Our working hypothesis, based on our data, is that there are two levels at which the task is performed: (1) A lower level at which state-action values are learned within the attended-to sub-state space. This sub-space corresponds to the current attention focus, where strategies map to distinct low-dimensional state spaces via their attention focus. (2) A higher level at which strategies are either selected from a pre-existing repertoire of hypothetical rules based on their value or created anew. Strategies are abstract because they summarize different combinations of environmental cues, actions, and potential outcomes within a common concept (in contrast to simple motor actions or cue-driven action selection), and this facilitates the transfer of learned skills to new tasks. Both computational layers contribute to strong dimensionality reduction during strategy-based learning. We model the lower level using attention-modulated RL models, and predict the existence of the higher level based on observed transfer effects (value-based strategy selection) and neural decoding analyses (abstract prefrontal representations). Despite differences in learning speed across species (and additional human-specific computational processes16), our results indicate that these computational mechanisms are shared with humans. Created in BioRender. Böhme, N. (2025) https://BioRender.com/wex0oga.
