Johns Hopkins University
Homewood Campus
(410-516-5250/office phone)

COGNITIVE SCIENCE DEPARTMENT

Thursday, March 5, 2009
3:45 PM

Dr. Nikolaus Kriegeskorte

NIH / NIMH

The transformation of representational similarity along human ventral-stream stages of visual object processing 

The human ventral stream is known to host high-level representations of visual objects, from which category information can easily be read out with linear decoders [1,2]. Using the method of representational similarity analysis [3, http://frontiersin.org/systemsneuroscience/paper/10.3389/neuro.06/004.2008/], we have previously shown that single-object-image response patterns in human inferior temporal (IT) cortex, when grouped by similarity, reflect conventional object categories [4] and that the categorical structure as well as the within-category similarity structure matches between human and monkey IT [4,5]. Here we start at early visual cortex and follow the ventral stream through key functional regions in order to understand how categoricality emerges across stages of processing.

Early visual cortex (EVC) exhibits a representational similarity reflecting visual shape (predicted, for example, by the dissimilarities of the silhouette images capturing the outer boundary of the objects). Categoricality (in the sense of categorical clustering of fMRI response patterns) first emerges at the level of the lateral occipital complex (LOC). The major categorical division is between animate and inanimate objects. Each region appears to exhibit a unique representational similarity, whose features are consistent with previous findings while challenging their prevalent interpretation. For example, the fusiform face area (FFA) exhibits a particularly tight cluster of response patterns corresponding to faces (including human and animal faces). However, FFA distinguishes animate from inanimate objects and even particular inanimate objects from each other better than it distinguishes individual faces. IT regions (including LOC and higher regions) appear to combine categorical and continuous representations. Within category clusters, IT represents object exemplars in a continuous object space, which may reflect a form of visual similarity.

References
[1] Kanwisher N et al. (1997) J Neurosci 17: 4302-11.
[2] Haxby JV et al. (2001) Science 293: 2425-30.
[3] Kriegeskorte N et al. (2008) Frontiers in Systems Neuroscience.
[4] Kriegeskorte et al. (2008) Neuron 60(6): 1126-41.
[5] Kiani R et al. (2007) J Neurophysiol 97(6), 4296-309.