articlesThe capacity to categorize stimuli is fundamental to all living organisms 1,2 . Theories of categorization agree upon the importance of the similarity between stimuli to account for many aspects of categorization performance [3][4][5] . However, it is not straightforward to compute the degree of similarity between stimuli that can vary across a high number of dimensions, like complex shapes. Fortunately, the similarities among a set of complex stimuli can often be described in a more compact way [6][7][8] . Indeed, stimuli from many behaviorally relevant sets can be represented in a low-dimensional representation space in which the proximity between stimuli is related to their similarity. For example, by presenting the randomly ordered shapes of Fig. 1d in a particular order (Fig. 1a-c), the similarities can be easily described by a twodimensional square-like configuration. Several behavioral studies that have varied complex shape differences parametrically revealed that primates are able to represent the similarities between shapes in a low-dimensional representation space without ever seeing these stimuli in their parametric configuration 9-12 .Here we aim to study directly the neural basis of these lowdimensional representation spaces. Object recognition and categorization in macaques is thought to depend on the inferotemporal cortex (IT) 13,14 . Single IT neurons are selective for moderately complex object features 15 , but several studies have found little relationship between the similarities between complex objects and the responses of single IT neurons 16,17 . However, one needs to manipulate shape similarity parametrically to investigate how the responses of IT neurons to complex stimuli are related to the proximity of these stimuli in a low-dimensional space. Thus, we investigated whether the response pattern across a population of IT neurons can reveal a low-dimensional and faithful representation of shape similarity using parameterized shapes. Behavioral studies with parameterized shapes have shown that the similarities among these complex stimuli can be represented using a low number of dimensions. Using psychophysical measurements and single-cell recordings in macaque inferotemporal (IT) cortex, we found an agreement between low-dimensional parametric configurations of shapes and the representation of shape similarity at the behavioral and neuronal level. The shape configurations, computed from both the perceived and neuron-based similarities, revealed a low number of dimensions and contained the same stimulus order as the parametric configurations. However, at a metric level, the behavioral and neural representations deviated consistently from the parametric configurations. These findings suggest an ordinally faithful but metrically biased representation of shape similarity in IT.As the analysis of the visual input in the visual system is highly nonlinear, the neuronal representation space could deviate from the configurations in parameter space in several ways. Previous psychophysical stu...