Effects of categorical and numerical feedback on category learning

Real-world learning signals often come in the form of a continuous range of rewards or punishments, such as receiving more or less money or other reward. However, in laboratory studies, feedback used to examine how humans learn new categories has almost invariably been categorical in nature (i.e. Correct/Incorrect, or A/Not-A). Whether numerical or categorical feedback leads to better learning is an open question. One possibility is that numerical feedback could give more fine-grained information about a category. Alternatively, categorical feedback is more dichotomous, potentially leading to larger error signals. Here we test how feedback impacts category learning by having participants learn to categorize novel line stimuli from either numerical, categorical, or a combination of both types of feedback. Performance was better for categorical relative to the more variable numerical feedback. However, participants were able to learn to effectively categorize from numerical feedback, and providing larger numerical rewards for easier, more representative stimuli was more successful in promoting learning than providing larger rewards for harder to classify stimuli. Simulations and fits of a connectionist model to participants' performance data suggest that categorical feedback promotes better learning by eliciting larger prediction errors than numerical feedback.