Modeling object representations in natural vision

Object perception relies on intensive processing in the humanoccitotemporal cortex (OTC). While its large-scale pattern ofobject selectivity has been widely described, looking at thecomputational processes controlling the spatial organization ofOTC has proven challenging, since visual dimensions aremutually correlated (e.g., object shape and identity). In thepresent thesis, we investigated how different object propertiesthat are relevant to behavior and share common variance arerepresented in our visual cortex in natural vision.In Chapter 1, we described how our exceptionally reliablevisual system transforms continuous retinal signaling intomeaningful objects. In addition, we demonstrated how thisprocess is challenging and complex, given the unreliability of theretinal input and the widespread mutual correlations betweenbehaviorally relevant object properties (e.g., shape and semanticcategory).In Chapter 2, we described an fMRI study on scenesegmentation. In this first study, we analyzed brain responsesduring passive natural image viewing. Subjects attended tohundreds of natural scenes and we derived brain representationsfrom each occipital region and compared them to parametricrepresentations, so to reveal the inner filtering operated by eachbrain region.In contrast to strictly hierarchical and compartmentalizedviews on brain selectivity, the whole occipital lobe is involved inthe high-level cognitive task of segmenting foreground andbackground, as early as V1. At the same time, contrast andspatial frequencies are represented also in higher visual regionssuch as V4 and LOC.However, due to the low temporal resolution of fMRI, thefirst study alone cannot resolve if those shared representationsreflect a common spatiotemporal process. Thus, in a secondMEG study, presented in Chapter 3, we derived brainrepresentations in time and space of subjects attending to different objects. We compared these representations to modelderivedrepresentations comprising V1-like features, objectshape and semantic category. We also employed a statisticalapproach that compute the relative weight (i.e., orthogonalcomponent) of each model in explaining the MEGrepresentations.By doing so, we found that a small cluster of posteriorsensors independently processes all the tested features as earlyas 100-150ms after stimulus onset. Thus, the same features can beretrieved in the activity of multiple regions, and orthogonalcomponents of those features are processed by the same corticalstructures at the same latencies. Thus, is there a broaderorganization determining these observations? What is the linkbetween coding of mutual and orthogonal objectrepresentations?In our third study, presented in Chapter 4, we employedfMRI to explore the spatial organization of sensitivity to mutualand orthogonal representations in the human visual cortex.Subjects attended to object pictures while performing anunrelated attentive task. By employing a variance partitioningmethod, we found that the weight of mutual representationsincreases along the visual hierarchy, from posterior to anteriorregions.Overall, these results depict a complex picture of our visualcortex. First, there is not a clear selectivity hierarchy, butinformation spreads between regions: early cortical areas accessto high-level representations and are involved in complexcognitive tasks (i.e., object segmentation). Second, concurrentprocessing of orthogonal object shape, contrast and categoryrepresentations is fast (100-150ms) in the right posterior brain.And, third, the visual cortex encodes mutual relations betweendifferent features in a topographic fashion while object shape isencoded along different dimensions, each representingorthogonal features.