Variance predicts salience in central sensory processing

Abstract

Information processing in the sensory periphery is shaped by natural stimulus statistics. In the periphery, a transmission bottleneck constrains performance; thus efficient coding implies that natural signal components with a predictably wider range should be compressed. In a different regime—when sampling limitations constrain performance—efficient coding implies that more resources should be allocated to informative features that are more variable. We propose that this regime is relevant for sensory cortex when it extracts complex features from limited numbers of sensory samples. To test this prediction, we use central visual processing as a model: we show that visual sensitivity for local multi-point spatial correlations, described by dozens of independently-measured parameters, can be quantitatively predicted from the structure of natural images. This suggests that efficient coding applies centrally, where it extends to higher-order sensory features and operates in a regime in which sensitivity increases with feature variability.DOI: http://dx.doi.org/10.7554/eLife.03722.001

Systematic differences between visually-relevant global and local image statistics of brain MRI and natural scenes

“…http://dx.doi.org/10.1101/271338 doi: bioRxiv preprint first posted online Feb. 25, 2018; the goals of this study --their distribution in natural images has been well-characterized [7], as has their salience to human observers [8]. The close match of these sets of findings [7] suggests that the human visual system is tuned to make efficient use of these natural image statistics, and motivated us to determine whether a similar match was present for MRI images.…”

“…The larger ( , ) R N parameter values used by Hermundstad [7] for natural images ( 64 R N   ) could not be used, because of the limited number of voxels per plane in the MRI images.…”

“…Then, as in [7], we whitened these ROI's by filtering them in the frequency domain, attenuating each Fourier component by the inverse square-root of the power spectrum. Then, these images were binarized at the median point, where the median was determined from all of the ROI's drawn from the same brain.…”

“…This matching is at a surprising degree of detail -including not only the well-recognized stage of redundancy reduction by removal of global correlations [9], [10], but also an analysis of local statistics in a way that is closely matched to the features that distinguish patches of natural images [7]. While the computational strategies in human vision are sufficiently robust and general to enable perceptual judgments about images that are highly non-natural -for example, modern art -there is ample evidence that these strategies reflect a specific allocation of computational resources: for example, some kinds of local correlations are readily perceptible, while others, which are of comparable mathematical complexity, escape our notice [11].…”

“…MRI image formation is also radically different from natural image formation, involving acquisition in Fourier space, followed by several filtering and processing steps that all have the potential to alter image statistics. As a consequence, aspects of the human visual system that are finely-tuned to processing natural scenes may not be equally well-suited for MRI images, and identification of the differences between MRI images and natural scenes may lead to improved machine analysis of MRI.With these considerations in mind, we analyzed spectra and local image statistics of MRI images in several databases including T1 and FLAIR sequence types and of simulated MRI images [1]- [6], and compared this analysis to a parallel analysis of natural images [7] and visual sensitivity [7][8]. We found substantial differences between the statistical features of MRI images and natural images.…”