Modeling Functional Map Coverage in the Primary Visual Cortex
Lars Crawford, Neurobiology
Bair Lab, University of Washington 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 0 1000 2000 3000 4000 5000 6000 7000 Spatial Frequency (cycles/degree) Cell Count 1 2 3 4 5 6 7 8 0 1000 2000 3000 4000 5000 6000 7000 8000 Spatial Frequency (cycles/degree) Cell Count 0.5 0.7 1.0 1.4 2.0 2.8 4.0 5.6 8.0 0 1 2 3 4 5 6 0 500 1000 1500 2000 2500 3000 Spatial Frequency (cycles/degree) Cell Count 0 1 2 3 4 5 6 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Spatial Frequency (cycles/degree) Cell Count 1 2 3 4 5 6 7 8 0 1000 2000 3000 4000 5000 6000 7000 Spatial Frequency (cycles/degree) Cell Count 0.5 0.7 1.0 1.4 2.0 2.8 4.0 5.6 8.0 10 Randomized Warps Orientation (Ori) Spatial Frequency (SF) Ocular Dominance (OD) Map Warping: - Functional map layout within these models emulate biology with respect to general shape: orientation pinwheels, spatial frequency blobs and ocular dominance columns. - True biological layout of these maps is far more noisy. - A series of two dimensional radial Gabor spatial filters were applied to add noise to the maps in an attempt to make them appear more biologically plausible. Coverage Factor vs. Nwarps: - C values are far lower than in the literature due to exquisitely uniform maps. - To asses the effect of map rigidity vs. noisyness on coverage, c values were calculated for maps with varying numbers of Gabor spatial filters. - Noise moves maps away from uniformity. Coverage Factor & Activity Maps: - Activity maps were constructed and coverage factor, c, calculated for several models with varying distributions of spatial frequency preference. These activity maps represent stimulus space. - Taken together, c' and the associated map of neuronal activity in stimulus space reveal the coverage uniformity of map variable combinations. Introduction: The brain effortlessly carries out the daunting task of visual perception by utilizing millions of neurons to process various aspects of the visual scene. The systematic arrangement of these cells across the primary visual cortex (V1) can be described by a set of superimposed maps that convey the orientation, scale, location and eye of origin of the image features that best excite the neurons. Current theory suggests these maps are arranged such that all combinations of stimulus parameters are covered equally and continuously in local regions across the visual cortex. Here we develop a model that attempts to reconcile the idealization of even coverage with known physiological and anatomical data. Acknowledgments: I would like to thank Wyeth Bair for being an excellent mentor as well as Pamela Baker, both for their guidance through my research from coding to model construction to data analysis. This research was funded by the NIH Computational Neuroscience Training Grant 5R90DA033461-03, and NSF CRCNS grant IIS-1309725. Example V1 Receptive Field - V1 receptive fields: multiple LGN ON/OFF cells arranged to produce preference for spatial frequency, orientation, eye. - 8 Ori x 25 (x,y) x 3 SF x 2 eyes = 1200 unique stimuli. - V1 cells recieve conductance inputs from leaky integrate- and-fire LGN cells with ON/OFF center surround receptive fields. ON cell OFF cell Methods: - We have conducted computer simulations to measure the evenness of coverage in large-scale, biologically plausible models of V1 composed of 224x112 cells arranged in functional maps. Functional Map Layout Ocular Dominance Map (OD) Left Eye Right Eye Orientation Map (Ori) Spatial Frequency Map (SF) Low Mid High ON OFF 1 Degree Visual Angle Ori SF Idealized V1 Receptive Field Example Stimulus 1 Ori - 112.5 deg 1 SF - 1 cyc/deg 1 (x,y) Location - (2,2) 1 Eye - left Cortical Activity Response to Example Stimulus 1 2 3 4 5 6 7 8 0 2000 4000 6000 8000 10000 12000 14000 Spatial Frequency (cycles/degree) Cell Count 0.5 0.7 1.0 1.4 2.0 2.8 4.0 5.6 8.0 0 1 2 3 4 5 6 0 1000 2000 3000 4000 5000 6000 Spatial Frequency (cycles/degree) Cell Count 1 2 3 4 5 0 5000 10000 15000 0.5 0.7 1.0 1.4 2.0 2.8 Spatial Frequency (cycles/degree) Cell Count C = / (over all pixels [all stimuli] in a particular map) Activity Map Pixel (1 unique stimulus) = ( individual V1 cell activity) / n cells Unwarp Maps Orientation (Ori) Spatial Frequency (SF) Ocular Dominance (OD) Ori Tuning Curve 50 0 50 100 150 200 0 10 20 30 40 50 60 70 80 90 Orientation (deg) R e s p o n s e (s p k / s ) 0 1 2 3 4 5 6 7 8 9 0 10 20 30 40 50 60 70 80 SF Tuning Curves Low Mid High Spatial Frequency (cyc/deg) R e s p o n s e (s p k / s ) 0 10 20 30 40 50 0.041 0.042 0.043 0.044 0.045 0.046 0.047 0.048 0.049 Number of Warps C o v e r a g e F a c t o r (c ) Discussion: The concept of coverage uniformity is appealing because it corresponds to the intuition that we should be able to see the features of a simple shape equally well at any rotation and at a variety of distances, and it is important because it places a constraint on the architecture of the cortex. The data presented here indicate that uniformity requires the cortex have a SF distribution skewed drastically toward high SF preference to account for a tiling effect (similar to how many small tiles are required to fill the space of a large tile). However, that distributions of SF found in the literature are not comparable to those used here suggests two ideas: 1) the coverage constraint is not met in the primate primary visual cortex or 2) high SF is somehow being occluded during cellular tuning experiments. Further imaging studies are necessary to illucidate the true SF distribution in V1 in order to more accurately determine the uniformity of coverage with respect to stimulus features. 1.3 1.4 1.5 1.6 1.7 1.8 = 1.723595; = 0.044434; c = 0.025780 High SF Skew Model A v e r a g e d
A c t i v i t y
( s p k s / s ) Most uniform map has SF distribution skewed toward high SF. 1.3 1.4 1.5 1.6 1.7 1.8 = 1.424482; = 0.058886; c = 0.41339 Low SF Skew Model A v e r a g e d
A c t i v i t y
( s p k s / s ) 1.3 1.4 1.5 1.6 1.7 1.8 Linear SF Transition Model = 1.544816; = 0.067551; c = 0.043727 A v e r a g e d
A c t i v i t y
( s p k s / s ) L R Low Mid High 2 x OD 3 x SF 8x Ori 1.3 1.4 1.5 1.6 1.7 1.8 = 1.538703; = 0.048843; c = 0.031743 Even SF Distribution Model X Location: 1 2 3 4 5 Y Location: 1 2 3 4 5 A v e r a g e d