Abstract
A growing body of neuroscience research focuses on linking computational models and measurements of neural activity. One such line of research centers on visual face recognition. A number of neuroimaging studies have reported mirror-symmetric tuning in human face-selective areas, in line with single-neuron observations in the macaque anterior-lateral face patch. Overall, however, studies have reported conflicting results regarding the form of viewpoint tuning. We have previously argued that these inconsistencies reflect differences in data analysis methods as well as low-level stimulus confounds. To date, no computational model has been able to account for the observed similarities and differences between these studies. Here, we propose a simple computational model that includes a small number of biologically-interpretable parameters. Our model replicates a number of key results reported across studies investigating viewpoint sensitivity in human face-selective areas. We implemented a multilayered, randomly-connected feedforward network incorporating prominent biological constraints including: (i) cortical magnification in the input layer, (ii) incremental interhemispheric crossing of left and right hemifield representations in subsequent layers, and (iii) density of network connections. We also (iv) parameterized signal-to-noise characteristics, and (v) tested the impact of static measurement gain fields on the similarity structure of simulated brain patterns across different network layers. Our model takes as input the same face stimuli used in some of these neuroimaging studies, and accounts for a number of commonalities as well as important inconsistencies among studies. A single aspect of our model, the gradual increase of interhemispheric connections across layers, was sufficient to replicate view-tuned representations in early stages of the visual hierarchy, as well as a prevalence of “mirror-symmetry” in later processing stages. Differences in data analysis procedures accounted for the remaining inconsistencies observed in previous studies. Our results underscore the importance of incorporating biological constraints in computational models that aim to explain empirical observations.