Bruno Google Lecture 2005

Mar 16, 2011 01:01 PM

a wonderful overview of vision and the challenges.  delivered at Google in 2005.

Starts with the evolution of the eye  (cites Michael Land and Russell Fernald work)

see node on evol of eye... wonderful stuff.

Rhodopsin in billions of years old and is used by essen all seeing creatures...

retinal + difft proteins.

pit eye then pin hole eye then refrac material for  lens (to allow more light 

and yet maintain focus),  mirrors at back,  spherical lenses in fish.

how about brain itself:  we are in the pre-Newtonian era in terms of understnading this.

FLY  compound eye  (gary larsen cartton on lady with fly swatter)

each eye segment has only 7 photoreceptors  per facet....  

in insects/ arthropods  eyes are attached to exoskeleton... cannot move...

have to move whole body.

fly's  turn at 3,000 degrees per second while flying...  mate in the air  etc.

eye speeds are super fast.

jumping spiders... stalk prey like a cat.

rods are max sensitive btwn blue and green.

cones  sensitive at  red, blue and green.

(blue is out of focus...  not may of those blue cones)

slide with mosaic of cones... most are red cones, some patches of green... uncommon blue.

head is constant wobbling;  eyes wobbling  BUT

eye mvmt partially compensates for head motion.

BRUNO  says   120 cones per degree in fovea   (fewer than my estimate of 200).

its a complete mystery how we get a stable visual field with all that motion.

(ie despite cone arrray dancing around).

photoreceptors first go to horizontal cells  (do lateral inhibition to sharpen resolution)

then to bipolar cells  (relay info)

then to amacrine and finally to ret. ganglion cells.

fovea:  1 to 1 with retin ganglion cells;   There are  1M to 1.5 M  ret gangl cells per eye (in each optic nerve)

He loves the book by Robt Rodeick  (see node)  now deceased, Prof at Univ of Wash

The First Steps in Seeing   (unfort v expensive).

for the above, see my node on rods and cones

as you bring your hand in toward your face... you get no added info cuz

the resolution drops off peripherally in your retina.

the retinal ganglion cells are gathering impulses from many receptors at the periphery....

taking an average.

there are 9X as many PARVO gangl cells as there are MAGNO cells.

(the parvos  are the one to ones)

1M to 1.5 M  fibers in each optic nerve.

LGN is 1 to 1 with ret ganglion cells  (parvos and magnos are handled in separate layers)

the LGN is surrounded by an inhibitory layer. called the reticular nucleus.

There are 10X the number of fibers going BACK from cortex to LGN as from retina to LGN.

he quotes David Mumford's theory  (cortex is sending a skeleton of what it thinks is going on in visual

world to LGN.

Marty Usney   can find exact cell that a partic retin gangl cell stimulates in LGN and in V1

to get exact correspondences.

LGN field is either on center or off center (same as gangl cell)

V1 pattern is a Gabor  with on or off center  (but that is really only just a minority of cells, like 15% of them)

What are the other 85% of neurons in V1 doing????

V1 is NOT just a big gabor filter.

Macaques  (pron muh KACKS)   v1, v2 and v4 are all retinotopic.

KEY slide:   1 mm2 of V1  is say about 2mm thick and

has about 100,000 neurons per mm3  but

it gets only about 800 parvo axons from LGN.

that is 400 from EACH EYE ;  divide that by TWO (a pair of one each of ON cell and OFF cell)

so get 200 nodes or pixels per mm2 of V1 =  or about a 14 by 14 pixel array from retina

So,  200 pixels are being processed by hundreds of thousands of neurons.

Question:  what is that circuitry doing?

The STANDARD MODEL (the hubel and wiesel model) is that gabor is convolved with the input.

but it is not this simple

most neurons are NOT firing.

he discusses 5 biases in gathering V1 (or other data)

1) the neurosci  only gather data from neurons that are briskly firing... ignore the others

2)  they use only biased stimuli...  just looking for what works really well

30 artif testing conditions... not natural vision.  etc.

mean firing rate in mammals of neurons is about one hertz... once per second.

so, standard model is a biased theory... self-fulfilling prophecy.

again, look at V1....  only 5% of input is from LGN...  the other 95% is from other cortical neurons.