Abstract
Both shape and color can be used to make inferences about object identity, but the mechanisms that integrate these features are not well understood. Humans learn about shapes faster than colors; moreover, knowledge about typical colors associated with objects is thought to influence color appearance (“cognitive penetrance”), suggesting both independent and combined processing. We tested whether these behaviors generalize to macaque monkeys using a reinforcement-learning framework applied to behavioral data collected in three animals over 5 years. In the “training” phase, monkeys were shown one of fourteen colored shapes (“objects”), and then rewarded for selecting the matching colored blob or achromatic shape from several choices. The objects’ color-shape combinations were fixed. After reaching plateau performance (mean=96.8%), they began the “probe” phase: the cue was either a colored blob or achromatic shape of one of the original fourteen objects, and monkeys were rewarded for selecting the corresponding other feature (achromatic shape or colored blob) learned in the training phase. Data were fit using reinforcement-learning models. In the training phase, animals learned to associate the object to its shape faster than to its color (mean rate=0.011 vs. 0.0075: p<0.01). In the probe phase, animals neither started above chance performance nor transferred information from training, suggesting they did not benefit from cognitive penetrance (both tests: p>0.05). Over the following probe sessions, rewarded shape-to-color associations were learned faster than color-to-shape associations (mean rate=0.0022 vs 0.0016: p<0.01). The asymmetry in learning about shape (faster) versus color (slower) in both training and probe phases matches predictions from human development. But the results contradict predictions that over-learned color-shape associations automatically cause achromatic objects to appear tinted with their typical colors, arguing against cognitive penetrance in the task. Finally, the reinforcement-learning model’s success in capturing these results suggests investigating color-shape learning will uncover general principles of reward-based learning.