Abstract
We have recently demonstrated the successful time-point-by-time-point decoding, with multivoxel pattern classification, of remembered directions of motion during a short-term delayed-recognition task (Riggall & Postle, 2012). Importantly, this item-level decoding was only possible within medial and lateral occipital cortex ROIs and not in frontal and parietal ROIs that contain robust, elevated delay-period signal. Here, we address the seemingly paradoxical fact that, in the same data, 'importance maps' from whole-brain results, for which decoding performance was slightly better than in any individual ROIs, contained a number of important voxels located in frontal and parietal regions. We used an iterative approach to determine the optimal set of whole-brain voxels necessary for maximal classification performance. Starting with the single-most predictive voxel, we iteratively added the next most predictive voxel until maximal classification performance was achieved ('build-up model'). We complemented this with the reverse process, starting with the full set of feature-selected voxels and iteratively removing the least informative voxels ('knock-out model'). These two approaches provided generally comparable results. Voxels outside of the medial and lateral occipital cortex were added late in the build-up model (always within the last 33% of voxels added) or dropped early in the knock-out model (within the first 40% of voxels dropped). Large-scale simulations of the data, varying the relative signal-to-noise ratio and correlations between patterns in different regions and repeating the build-up and knock-out classification analyses, suggest that such results are consistent with small contributions from non-occipital regions, but only when they are correlated with posterior regions. These results suggest that, whereas frontal and parietal regions may participate in broadly distributed representations of trial-specific information, these regions cannot be construed as storing 'independent' mnemonic representations.
Meeting abstract presented at VSS 2014