AbstractClassification, one of the key ingredients for human cognition, entails establishing a criterion that splits a given feature space into mutually exclusive subspaces. In classification tasks performed in daily life, however, a criterion is often not provided explicitly but instead needs to be guessed from past samples of a feature space. For example, we judge today’s temperature to be “cold” or “warm” by implicitly comparing it against a “typical” seasonal temperature. In such situations, establishing an optimal criterion is challenging for cognitive agents with bounded memory because it requires retrieving an entire set of past episodes with precision. As a computational account for how humans carry out this challenging operation, we developed a normative Bayesian model of classification (NBMC), in which Bayesian agents, whose working-memory precision decays as episodes elapse, continuously update their criterion as they perform a binary perceptual classification task on sequentially presented stimuli. We drew a set of specific implications regarding key properties of classification from the NBMC, and demonstrated the correspondence between the NBMC and human observers in classification behavior for each of those implications. Furthermore, in the functional magnetic resonance imaging responses acquired concurrently with behavioral data, we identified an ensemble of brain activities that coherently represent the latent variables, including the inferred criterion, of the NBMC. Given these results, we believe that the NBMC is worth being considered as a useful computational model that guides behavioral and neural studies on perceptual classification, especially for agents with bounded memory representation of past sensory events.Significant StatementAlthough classification—assigning events into mutually exclusive classes—requires a criterion, people often have to perform various classification tasks without explicit criteria. In such situations, forming a criterion based on past experience is quite challenging because people’s memory of past events deteriorates quickly over time. Here, we provided a computational model for how a memory-bounded yet normative agent infers the criterion from past episodes to maximally perform a binary perceptual classification task. This model successfully captured several key properties of human classification behavior, and the neural signals representing its latent variables were identified in the classifying human brains. By offering a rational account for memory-bonded agents’ classification, our model can guide future behavioral and neural studies on perceptual classification.
Filtering in feature space: a computational model of grouping by proximity and similarity