Purchase this article with an account.
Luca Vizioli, Lucy Petro, Lars Muckli; Decoding the spatial scale of information in visual cortex. Journal of Vision 2014;14(10):1087. doi: https://doi.org/10.1167/14.10.1087.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Functional brain imaging has a spatial resolution in the range of millimetres, too low to directly capture the columnar substructures of visual cortex. However, multi-voxel pattern analysis (MVPA) may recover columnar-grained, micro-level features such as orientation preference. This claim has recently been challenged by the demonstration that decoding of microscopic features within visual cortex was in fact driven by macroscopic scale organization information, which co-varies with the former (Boynton, 2005, Freeman et al., 2011). The question of whether MVPA describes fine- or coarse-grained information pattern is therefore greatly debated. Here we propose a simple data driven approach to estimate the spatial scale of MVPA. We systematically investigated the impact of (mis)alignment upon support vector machine (SVM) classification performance on both 3T and 7T feedforward (i.e. visually stimulated) and feedback (i.e. visually non-stimulated) signals in V1 elicited by natural scene stimuli (Smith and Muckli, 2010). We simulated different extents of misalignment across functional runs by parametrically shifting our region of interest (ROI) one to five voxels in both directions along all three axes. The artificially misaligned ROis were then grouped according to number of voxels shifted (ranging from 0 to 5), regardless of direction. We trained an SVM classifier on the original and tested it on the misaligned data, independently per misalignment percentage. Our results illustrate a significant drop in SVM classification performance as a function of misalignment. The SVM performance curve, best described by a non-linear model, was characterised by a steep decrease in classification performance for the initial shift (i.e. 1 voxel), attenuating over larger shifts. As well as being of interest to all researchers implementing an unsupervised learning algorithm across functional runs, these results indicate that multi-voxel patterns activity hold both fine as well as more coarse grained neural information.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only