[1] Sanborn, AN, Griffiths, TL. Markov chain Monte Carlo with people. NIPS, 2008. [2] MacKay, D. Information Theory, Inference, and Learning Algorithms. 2001.

Cortical responses were measured using fMRI (3T Siemens Allegra, 8-ch phased-array surface coil, 2 × 2 × 2 mm, 24 sl perpendicular to calcarine sulcus), while subjects viewed an oriented sinusoidal grating (0.5 cycles/deg) that filled a 4 deg peripheral annulus with a smooth edge. The orientation of the stimulus changed every 1.5 s, cycling through sixteen evenly spaced angles (0-180 deg). The response of each voxel was fit to a sinusoid with period of the stimulus. The phase of the best-fitting sinusoid indicated the preferred orientation of the voxel.

We observed a topographic map of orientation preference in human V1, confirming and extending previous reports of a quadrant bias for radial orientations (Sasaki et al, Neuron, 2006). The map was tightly co-localized with the retinotopic map: at each location within V1, responses exhibited a preference for radial orientations. Circular correlation was used to quantify the similarity between the orientation and polar angle maps (r = 0.75; p < 0.0001). Control experiments confirmed that the orientation map was robust to a variety of stimulus parameters, and was not due to either attention or eye movements. Multivariate classification analyses were applied to decode stimulus orientation. Averaging the data in a manner consistent with the structure of the topographic map did not affect decoding accuracy, demonstrating that the map was sufficient to classify orientation. Our results strongly suggest that orientation decoding does not reflect the irregular spatial arrangements of orientation columns.

There is a coarse-scale topographic map of orientation in V1. The orientation map provides a parsimonious explanation for how multivariate classification methods decode stimulus orientation from fMRI measurements, and challenges the conjecture that decoding reflects random irregularities in the fine-scale columnar architecture.

If this model accurately reflects representations in early extrastriate areas, then images synthesized to produce identical model responses should be metameric to a human observer. For each of several natural images and pooling region sizes, we generate multiple samples that are statistically-matched but otherwise as random as possible. We use a standard psychophysical task to measure observers’ ability to discriminate between image samples, as a function of the rate at which the statistical pooling regions grow with eccentricity. When image samples are statistically matched within small pooling regions, observers perform at chance (50%), failing to notice substantial differences in the periphery. When images are matched within larger pooling regions, discriminability approaches 100%. We fit the psychometric function to estimate the pooling region over which the observer estimates statistics. The result is consistent with receptive field sizes in macaque mid-ventral areas (particularly V2).

Our model also fully instantiates a recently proposed explanation (Balas et al., 2009) of the phenomenon of “visual crowding”, in which humans fail to recognize a peripheral target object surrounded by background clutter. In our model, crowding occurs because multiple objects fall within the same pooling region and the model responses cannot uniquely identify the target object. We synthesize images that are metameric to classic crowding stimuli (e.g. groups of letters), and find that stimulus configurations that produce crowding yield synthesized images with jumbled, unidentifiable objects.

We measured fMRI responses in multiple retinotopic areas while subjects viewed texture samples drawn from 6 different texture categories. Categories were based on texture images selected from the Brodatz database. These images were preprocessed to ensure that their power spectra, averaged over orientation, were equal. For each category, many samples were generated using the Portilla/Simoncelli model. On each trial (inter-stimulus interval: 3.5-6.5 s), we showed a 1 s burst at 5 Hz of random texture samples drawn from the same category, mitigating any response specific to a particular texture sample. Texture category was randomized across trials.  Stimuli were presented within an annulus (radii: 0.5 and 9 deg). To divert and control attention, subjects performed a difficult letter-identification task at fixation.

We used a multivariate pattern-classification analysis with leave-one-run-out cross validation to test whether distributed patterns of fMRI responses reliably differed between the texture categories. Classification was significantly above chance in multiple visual areas, including primary visual cortex (V1), and areas V2 and V4. Classification performance was more robust across texture categories in V2 and V4 than in V1. Also, classification performance in V1 was highest for those texture categories with larger orientation differences, but this was not the case for V2 and V4. We conclude that responses in striate and extra-striate cortex reflect the statistical properties that distinguish natural texture categories, and that higher-tier areas extract increasingly complex statistics.

We used fMRI to measure the effect of letter crowding (which disrupts feature integration while preserving feature detection) on inter-area correlations in the ventral visual pathway. Observers viewed closely-spaced letters (8º eccentricity, presented at 1 Hz for blocks of 15-21 s, separated by 15-21 s blocks of no stimulation). Adjacent letters were displayed in alternation in the uncrowded condition and simultaneously in the crowded condition. During fMRI, observers performed a demanding contrast discrimination task at fixation to ensure that attention was diverted from the letter stimuli. In a separate psychophysics experiment, we confirmed that letter identification was impaired by crowding under these stimulus conditions. Retinotopic visual areas (including V1, V2, V3, V4) were defined in a separate session using standard procedures, and the visual word-form area (VWFA) in inferotemporal cortex was defined by measuring responses in alternation to English and Chinese character strings. Sub-regions of each retinotopic area were identified corresponding to the letter locations in the main experiment. To quantify interactions between cortical areas, we first removed the mean stimulus-driven response, separately for each cortical area and each condition (crowded, uncrowded), and then computed pairwise correlations between the residuals for each pair of cortical areas and both conditions. The residual correlations specifically reflected interactions between areas rather than common input from the stimulus.

Crowding reduced residual correlations between several pairs of visual areas, particularly between early visual areas (V1-V4) and VWFA. Differences in residual correlations occurred despite little or no differences between conditions (crowded, uncrowded) in either mean response or residual variance. Differences in residual correlations were eliminated when letters were replaced with Gabor stimuli, which are elementary features not requiring integration.

We conclude that interactions between visual areas, especially between early areas and higher inferotemporal areas, reflect feature integration during object processing.