Effects of prototype abstraction on pattern completion and inference in concept space

Introduction: The processing of complex environments is greatly facilitated by the formation and use of concepts. Concepts represent combinations of features shared by similar entities and allow generalisation from limited experience to novel situations. Recent research suggests that map-like codes in the hippocampal-entorhinal system can support concept learning by representing the relations between experiences along relevant feature dimensions (Constantinescu, 2016; Park et al., 2020; Theves et al., 2019; Theves et al., 2020). In a behavioral and functional magnetic resonance imaging (fMRI) experiment, we investigate if this map-like representation of concepts supports the retrieval of abstracted information to guide inference. Methods: 40 participants performed a novel concept learning paradigm inside the fMRI scanner (3T MRI, TR: 1500 ms, TE: 22 ms, voxel size: 2.5 mm). Participants were trained to categorize a set of visual exemplars based on the ratio of their two continuous features (see figure below). Subsequently, they encountered exemplars that exhibited only one of the features and were instructed to complete the missing feature in a continuous fashion according to their category label. This concept learning and inference test phase was preceded and followed by stimulus viewing blocks. We examined whether feature inferences were more attracted by the previously experienced exemplar locations as suggested by an exemplar learning account (Medin & Schaffer, 1978; Nosofsky, 1984) or by the category prototypes as suggested by an abstract representation account (Posner & Keele, 1968; Reed, 1972). To this end, we trained a support vector regression on multivariate fMRI responses from the stimulus viewing blocks to decode completion responses during the feature inference task. Results: Participants improved their categorization performance well above chance during training, suggesting that they were able to learn the category structure. Both categories were classified equally well. Subjective prototypes that were generated by the participants at the end of the study lied within their respective category regions in feature space. fMRI-based feature decoding revealed moderate to high correlations with true feature dimension values in the visual cortex. We found that feature completions were attracted more towards the prototype location than to the nearest experienced exemplars, suggesting the retrieval of an abstract representation during feature inference. Conclusions: Our findings are consistent with the hypothesis that abstract representations are formed during concept learning. Furthermore, our results indicate that abstracted representations are employed to infer the missing information of an incomplete pattern. This study provides a first investigation of the relation between abstract category representations and cognitive maps