Independence of luminance and contrast in natural scenes and in the early visual system

Valerio Mante, Robert A Frazor, Vincent Bonin, Wilson S Geisler, and Matteo Carandini

Nature Neuroscience dec2005

Independence of luminance and contrast in natural scenes and in the early visual system

Valerio Mante, Robert A Frazor, Vincent Bonin, Wilson S Geisler, and Matteo Carandini

Nature Neuroscience dec2005

• measured natural statistics of local luminance, contrast

• modeled changing temporal kernel in cat LGN cells

• results: luminance independent of contrast kernel is

separable, too

• implications?

statistics of natural scenes

simulated saccade sequence

luminance

contrast

weighted local patch

movements sampled from measured distributions (uniform gave same results)

statistics of natural scenes

large dynamic range

little correlation from fixation to fixation

statistics of natural scenes

statistics of natural scenes

statistics of natural scenes

statistics of natural scenes

what causes these distributions?

• 1/f statistics

• phase alignment

• natural scene structure: illumination, reflectance, areas of high-luminance/high-contrast

what are the implications for neural coding?

• large dynamic range requires adaptation

• expect independent coding of independent quantities

neural sensitivity to luminance/contrast

luminance: 56→32 cdmluminance: 32→56 cdm

linear prediction

neural sensitivity to luminance/contrast

luminance: 100→31%contrast: 31→100%

linear prediction

measured response at fixed luminance, contrast

spiking rate varies with temporal frequency, contrast, luminance

model of neural response

linear filtering by convolution with spatio-temporal kernel

additive noise

thresholding non-linearity

the spatio-temporal kernel

the spatio-temporal kernel

spatial components

the spatio-temporal kernel

spatial components

temporal kernel (impulse response)

fitted params:

fitting the temporal kernel

descriptive model

fit parameters for each luminance/contrast setting

fitting the temporal kernel

descriptive model

fit parameters for each luminance/contrast setting

model each temporal kernel as a convolution of contrast, luminance, and base kernel (product in the freq domain)

separable model

fitting the temporal kernel

descriptive model

fit parameters for each luminance/contrast setting

results - % variance of neural response explained

both kernels work equally well

separabledescriptive

results - adaptation effects modeled with separable kernel

circles: neural responselines: predictions of model

luminance = 10%

luminance = 84%

contrast = 10%

contrast = 100%

discussion

dynamic range, speed of adaptation

stimuli

• what about other non-linear response properties? (cross-orientation, surround suppresion, etc)

• separate underlying mechanisms?

• what about responses to more complex images?

relationship to normalization models?

what are the neural mechanisms?

what are the functional implications?

end