Poster Presentation

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

Interactions Between Features and Conjunctions during Category Learning

Heeseung Lee¹, John D Murray¹; ¹Dartmouth College

Presenter: Heeseung Lee

In learning multi-feature categories, humans often rely on primary features themselves rather than conjunction of these features. Conjunctions have also been recognized as critical for learning complex tasks, as conjunctions can expand the dimensionality of neural representations to allow for linear classification of categories which are nonlinear in the primary features (e.g. XOR). When humans learn a task requiring use of both feature and conjunctive information, it remains unclear how feature and conjunctive contributions to choice behavior interact across the learning trajectory. Here, we investigate the interactions between feature and conjunctive representations with a task in which humans learn categories from stimuli composed of fractal images at three spatially defined features. We develop a kernel-based framework to decompose behavioral choices into contributions from feature and conjunctive learning modes. The decomposed learning dynamics showed three intriguing interactions in learning of features and conjunctions within a task block: (i) anti-correlation between loadings of primary features; (ii) enhanced loading of a conjunction yoked to a highly loaded primary feature; and (iii) reduced loading of a conjunction not yoked to that primary feature. Furthermore, we tested how learning the relative ‘weights’ of different features transfers from one task block to the next block, which requires learning new categories from new fractals presented at the same spatial feature locations. High loading on a conjunction during the prior block led to enhanced learning of a yoked primary feature during the subsequent block. We found that multiple prominent category-learning models fail to capture all four interactions observed in our experiments. Finally, eye-tracking during task performance implicates spatial attentional mechanisms in mediating these feature interactions. Together, our findings reveal multiple forms of interaction between primary and conjunctive features during category learning, which poses constraints for future models of learning, and its interplay with attention, in multi-feature environments.

Topic Area: Memory, Learning & Knowledge Structures