cortical circuits for vision jamie mazer neurobiology of cortical systems lecture 7 march 12, 2012
TRANSCRIPT
Cortical circuits for vision
Jamie Mazer
Neurobiology of Cortical Systems
Lecture 7
March 12, 2012
Readings for Thursday
How much of cortex is visual? (in primates)
Van Essen flat mapof macaque cortex
Primates are likely an extremeexample or an upper bound..
How much of cortex is visual?
Van Essen flat mapof macaque cortex
“simplified” Felleman & Van Essen hierarchy
Key concepts
• phenomenon vs. implementation vs. function
• “centrally synthesized maps”– everything we perceive must be encoded by the retina
– if so, what’s all that visual cortex doing?
– generating explicit sensory representations
– “emergent” properties seem to be a key feature of high-level sensory cortical function
– Question: Is cortex required to generate explicit or abstract properties?
– Answer: What’s emergent in the retina? What about animals with not cortex, like birds and fish?
• are there common “motifs” across sensory modalities?– computational maps in other modalites?
– what about other species? are they unique to cortex?
Retinal bipolar cells receptive fields
Retinal ganglion cell RFs (only retinal output)
Receptive fields and center-surround opponency
Receptive fields and center-surround opponency
Center-surround organization Observed phenomenon? Implementation? Function?
Receptive fields and center-surround opponency
Center-surround organization Observed phenomenon? Characteristic RF structure Implementation? Lateral inhibition Function? Spatial derivative; contrast enhancement
Behavioral consequences of center surround organization
herring gridmach bands
Behavioral consequences of center surround organization
herring gridmach bands
Thalamus: dLGN
What changes between the photoreceptors and LGN?
• transition from receptor potentials to spiking• center-surround spatial receptive fields• “color opponency” (B-Y/R-G) instead of simple cone-
based wavelength tuning• segregation into parallel processing streams
– sustained and transient– fast and slow– on and off channels– color and luminance
Which brings us to primary visual cortex (BA 17; V1)
m
visualassociation
primaryvisual
Topographic organization of V1
- retinotopy- orientation columns- occular dominance columns- non-oriented blobs (L2)- orientation topography
Thalamocortical projections and the canonical microcircuit
Primary visual cortex: simple cell orientation tuning
hubel & wiesel 1968
orientation tuned V1 neuron
MOVIE
Primary visual cortex: simple cell orientation tuning
hubel & wiesel 1968
orientation tuned V1 neuronhubel & wiesel model
Primary visual cortex: simple cell orientation tuning
hubel & wiesel 1968
orientation tuned V1 neuronhubel & wiesel model
Key failures for the feedforward model?
- contrast invariant orientation tuning
Primary visual cortex: simple cell orientation tuning
hubel & wiesel 1968
orientation tuned V1 neuronhubel & wiesel model
Hubel & Wiesel Interpretation Observed phenomenon?
preferred orientation Implementation?
linear summation of LGN cells Function?
feature detectors for edges
Primary visual cortex: simple cell orientation tuning
hubel & wiesel 1968
orientation tuned V1 neuron hubel & wiesel model
Spatial Vision Interpretation Observed phenomenon?
preferred orientation Implementation?
quasi-linear combination of LGN cells Function?
spatiotemporal filtering
• cells prefer light increments or decrements
• cells have orientation tuning
• cells have a width tuning
• cells have length tuning
• cells have speed tuning
• cells are feature detectors where the feature is a bar of a particular orientation, size and speed
• intuitively obvious, simple to understand, seems to imply obvious behavioral function
• cells prefer light increments or decrements
• cells have orientation tuning
• cells have spatial frequency tuning
• cells have temporal frequency tuning
• cells are half-wave rectified spatiotemporal filters (Gabors)
• requires some math chops to understand, but has predictive power
Feature detector model “spatial vision” model
Primary visual cortex: spatial frequency tuning
Robson, DeValois, Maffei etc..
Feature detector model “spatial vision” model
• cells prefer light increments or decrements
• cells have orientation tuning
• cells have a width tuning
• cells have length tuning
• cells have speed tuning
• cells are feature detectors where the feature is a bar of a particular orientation, size and speed
• intuitively obvious, simple to understand, seems to imply obvious behavioral function
• cells prefer light increments or decrements
• cells have orientation tuning
• cells have spatial frequency tuning
• cells have temporal frequency tuning
• cells are half-wave rectified spatiotemporal filters
• requires some math chops to understand, but has predictive power
Primary visual cortex: simple complex
hubel & wiesel 1968
simple
complex
Primary visual cortex: simple complex
hubel & wiesel 1968
simple
complex
MOVIE
Primary visual cortex: simple complex
hubel & wiesel 1968
simple
complex
hypercomplex+length tuning+length tuning+length tuning
Primary visual cortex: simple complex
hubel & wiesel 1968
“simple cells”pool center-surround neuronsto form orientation selectivity
“complex cells”pool simple cells to becomeposition or phase invariant.
and turtles all the way down…
Complex cells and the F1/F0 ratio
cats
monkeys
Skottun et al, 1991
What’s the spatial vision model got to say?
Complex cells and the F1/F0 ratio
Skottun et al, 1991
cats
monkeys
Mechler & Ringach, 2002
is this all an artifact?
Reverse correlation and the spike triggered average
Jones & Palmer, 1987
Reverse correlation and the spike triggered average
Jones & Palmer, 1987
Reverse correlation and the spike triggered average
Jones & Palmer, 1987
V1 neurons are Gabor’s and Gabor’s are optimal…
Daugman, 1985
V1 neurons are Gabor’s and Gabor’s are optimal…
Daugman, 1985
Where do Gabor’s come from and the efficient coding hypothesis
Barlow, 1972
Where do Gabor’s come from and the efficient coding hypothesis
Where do Gabor’s come from and the efficient coding hypothesis
Vinje & Gallant, 2000
Where do Gabor’s come from and the efficient coding hypothesis
Haider et al, 2010
What have we established?• simple cells
– simple cells are partially assembled from LGN afferents
– one basic flavor: Gabor
• they are bar-detectors as well (glass half empty), but
• the Gabor-model seems like a more compact framework
• complex cells– complex cells are assembled from simple cells
– strict dichotomy not likely, more likely is,
• thalamocortical direct recipient simple cells, and,
• cells that are a combination of simple and non-simple innputs
• coding in V1– sparseness is a hallmark of an efficient code
– simple cells can be learned by maximizing sparseness
– sparseness in V1 is based on center-surround (intracortical) inhibitory interations
– the neural representation is awful close to what the computer vision people call a wavelet or multiscale pyramid and is the basis for things like MPG and JPG compression…
• perhaps we need more data from more complex stimuli?
Reverse correlation, complex cells and natural scenes
Reverse correlation, complex cells and natural scenes
Problems:1. STA doesn’t really work for
natural (non-white) stimuli2. the STA is just plain “wrong”
for complex cells
Linear receptive field maps in early vision
DeAngelis et al, 1995
still orientation tuned!where’s it coming from?
Reverse correlation, complex cells and natural scenes
Problems:1. STA doesn’t really work for
natural (non-white) stimuli2. the STA is just plain “wrong”
for complex cells
Reverse correlation, complex cells and natural scenes
Problems:1. STA doesn’t really work for
natural (non-white) stimuli2. the STA is just plain “wrong”
for complex cells
Spike Triggered Covariance (STC)
What have we established?• simple cells
– simple cells are partially assembled from LGN afferents
– one basic flavor: Gabor
• they are bar-detectors as well (glass half empty), but
• the Gabor-model seems like a more compact framework
• complex cells– complex cells are assembled from simple cells
– strict dichotomy not likely, more likely is,
• thalamocortical direct recipient simple cells, and,
• cells that are a combination of simple and non-simple innputs
• coding in V1– sparseness is a hallmark of an efficient code
– simple cells can be learned by maximizing sparseness
– sparseness in V1 is based on center-surround (intracortical) inhibitory interations
– the neural representation is awful close to what the computer vision people call a wavelet or multiscale pyramid and is the basis for things like MPG and JPG compression…
• perhaps we need more data from more complex stimuli?– STRFs, STC, regression analysis, MII etc all provide new tools could complex cells and complex stimuli…
• what did I not talk about??
Direction selectivity
hubel&wiesel 1968
MOVIE
Direction selectivity is Gabor-ish too (vs. Reichardt Detector)
DeAngelis et al, 1993. 1995
Disparity/depth tuning
focal plane
near
far
foveae G. Poggio et al
MOVIE
Disparity too fits in the spatial vision view…
Note: Complex cells see anti-correlated bars differently than correlated, not true for perception…
Ohzawa et al, 1997
Is spatial vision everything?
• the high-dimensional Gabor filter model explains a lot of the neurophysiological and psychophysical data, but..– finding the right dimensions is non-trivial as we’ll see next week.
– even when the dimensions are likely identified, it’s essentially a linear or quasi-linear model and doesn’t explain a range of observed phenomena, even in V1…
Center-surround interactions in V1 – generally NOT accounted for by the standard spatial vision model.
• end-stopping, length-tuning, “hypercomplexity” (H&W)• cross-orientation inhibition (Silito et al)• divisive gain control (Carandini paper!)• curvature processing (Dobbins & Zucker)• target pop-out (Knierim & Van Essen)• attention and/or figure segmentation (Lamme et al)
What are the emergent properties of V1?
• new features extracted– orientation– binocular disparity (depth)– direction selectivity– spatial frequency– color (really “transformed”)
• new maps– orientation (columns)– ocular dominance– segregation of color info in blobs
What have we established?• simple cells
– simple cells are partially assembled from LGN afferents
– one basic flavor: Gabor
• they are bar-detectors as well (glass half empty), but
• the Gabor-model seems like a more compact framework
• complex cells– complex cells are assembled from simple cells
– strict dichotomy not likely, more likely is,
• thalamocortical direct recipient simple cells
• cells that are a combination of simple and non-simple innputs
• coding in V1– sparseness is a hallmark of an efficient code
– simple cells can be learned by maximizing sparseness
– sparseness in V1 is based on center-surround (intracortical) inhibitory interations
– the neural representation is awful close to what the computer vision people call a wavelet or multiscale pyramid and is the basis for things like MPG and JPG compression…
• perhaps we need more data from more complex stimuli?– STRFs, STC, regression analysis, MII etc all provide new tools could complex cells and complex stimuli…
• V1 exhibits multiple emergent properties
• What happens when you lose V1??
• How much of this interpretation is primate-centric?
Primate-centric view?
Andermann et al., 2011
Neill & Stryker, 2010
Shuler & Bear, 2006
- Rodents have striate and extrastriate analogues or homologues- Tuning is similar, but not identical- “Extra-retinal” effects seem more pronounced
Readings for Thursday