Population responses in monkey IT, as measured with multiple single-unit recording, and fMRI response patterns in human VT cortex are related (Kiani et al., 2007 and Kriegeskorte et al., 2008b). Using our methods, the representational spaces for neuronal Selleck BMN673 population responses and fMRI response patterns could be modeled, preferably with data from the same animals, and the form of a transformation that relates the basis functions

for the neuronal space to the basis functions for the fMRI space could be investigated. The second goal of this project was to develop a single model that was valid across stimuli that evoke distinct patterns of response in VT cortex. To this end, we collected three data sets for deriving transformations into a common space and testing general validity. All data sets could be used to derive the parameters for hyperalignment, and all data sets allowed BSC of responses to different stimuli. The central

challenge was to estimate parameters in each subject for a high-dimensional transformation that captures the full variety of response patterns in VT cortex. We reasoned that achieving such general validity would require sampling a wide range of stimuli that reflect the Akt inhibitor statistics of normal visual experience. The use of a limited number of stimuli—eight, 12, or even 20 categories—constrains the number of dimensions that may be derived. We chose the full-length action movie as a varied, natural, and dynamic stimulus that can be viewed during an fMRI experiment (Hasson et al., 2004, Bartels and Zeki, 2004 and Sabuncu et al., 2010). Parameter estimates derived from responses to this stimulus produced a common model space that afforded highly accurate MVP classification for all three experiments. Supplemental analysis of the effect of

the number of movie time points used for model derivation indicates that maximal BSC required most of the movie (1,700 time points or 85 min; Figure S2D). This space has a dimensionality that cannot logically be derived from a more limited stimulus set. By contrast, the responses evoked by the stimuli in the category perception experiments did not have these properties. We also derived common models based on responses to the face and object categories in ten subjects Isotretinoin and on responses to the pictures of animals in 11 subjects. These alternative common models afforded high levels of accuracy for BSC of the stimulus categories used to derive the common space but did not generalize to BSC for the movie time segments. Thus, models based on hyperalignment of responses to a limited number of stimulus categories align only a small subspace within the representational space in VT cortex and are, therefore, inadequate as general models of that space. On the positive side, these results also show that hyperalignment can be used for BSC of an fMRI experiment without data from movie viewing.