B&B - Vision I
criticise available evidence
next steps + research
VisNet � primitive, what are the limitations
leila+ben: 4-5
us @6
---
questions
LGN: 6 layers (1-6 outside � inside). layers, like an onion. magno = the two innermost layers
no cross-talk between alternate layers of the different eyes. first place where the visual pathways come together for computation is in the cortex
why do the 2 pathways come together? stereopsis
no binocular neurons in the LGN (apart from trivial feedback effects) for depth
koniocellular streams � vanison article (K cells). unimportant � 3rd sort of visual stream, not much recognised
Cognitive Neursocicnces 1995 (Gazzaniga)
translation invariance � responding invariantly with respect to position, position independent
blobs. input to blobs goes straight through, not to layer 4 like everything else. where does it come from? K system slightly specialised for colour
usual view: everything comes in from the parvocellular system
cytochrome oxidase � enzyme, index of oxidative metabolism
that little bit of the cortex is more active than the nearby bits
the place where you find the blobs in the cortex
orientation columns (shift gradually across the cortex) suddenly show a singularity, where there�s no steady change
so not orientation specific
in all the columns, lateral inhibition � cells kept under control from neighbouring columns, except where you have a blob
just because you�ve got a break, you�ll have less inhibition, so will be more active, and so the need for cytochrome oxidase
may not be anything very special about the blobs at all � might be anomalies in the visual system network
high-D space � low-D space, you have to have anomalies � can�t be continuous everywhere
modularity
why is there segregation of function in the brain?
evolutionary advantage in having many systems
can modify each one a little bit at a time
modularity � why have a V4 and an MT
within the ventral stream, have modules for faces
parallel � speed
wiring length � modularity localises
V4 computation, as opp to MT computation
colour constancy, e.g. land�s retinex model
V4 on region � reddish light
at sunset, all the light goes red, but we still only see a certain amount of red
red on centre cell � white surround to subtract from the red
at sunset, the white light will appear reddish
want to calculate your best estimate of white by taking the average of all the other cells
sum together the outputs of all the other cells, and make them go into the surround
inhibitory neurons� average firing reflects the average luminant wavelength
subtract that from the value of the all the other neurons firing
doesn�t quite work
have a general mush driving the inhibitory neurons
if you mix in with all this computation, other neurons for global motion, will mean massive extra length of connections
= a crucial factor in brain weight. have as much white matter as grey matter. same problem with computers: interconnect, not transistors
so, length of wiring a major factor in producing segregation of function in the brain
also, simplification of wiring � genetics just need to tell �connect to your neighbour�, without having to specify which neighbour to avoid the motion/stereoscopic etc.
ontogeny = development to make you exist
means that you can have general rules
MT � could you compute colour constancy in V1
can�t get a big enough area (cf the aperture problem in motion)
as receptive fields get bigger, can do a certain class of problem that you couldn�t do before
MT = global motion
add up from a large number of cells (e.g. global motion of snow = downwards, even though individual particles will be moving in different directions), average their motion � can�t do that in a small receptive field
same argument for wiring length for segregation of function
modularity = nothing to do with brain damage, because wouldn�t be evolved (these other factors are far more general)
2 visual systems (or 3)
why not more?
2 classes of information to extract
statistics of the information itself
they�re not random like white noise � there�s a lot of structure
one sort of eigenvector (simple source of variance) is motion
if started with random information, might end up with 50D visual systems (if they all add to survival value)
�what� system � society?
certain bits of the brain will specialise for face and expression recognition
big bits (e.g. orbito-frontal cortex) for society = concerned with remembering the reward value of individuals
dorso-lateral prefrontal cortex � nothing to do with that � more for planning and STM
segregation starts at the ganglion cells in the retina
spatial frequency
advantage of analysing things in terms of spatial frequency: quantitative, allows linear analysis
break up a bar into individual components � split them up into different spatial frequency - do the neurons respond linearly to this
thought it would help with invariance
parvo are linear (white light over whole field, total sum = 0, so the gain is weighted towards the centre, roughly even/symmetric, though you do get hotspots)
simple vs complex � respond to the edge anywhere in their receptive field
assume then that complex cells come after simple, because need a larger receptive field
but complex cells fire before simple cells when electrically stimulated, so the synaptic routes may not be usual route/ecologically valid � not powerful evidence
and there are some direct LGN fibres � complex cells
hypercomplex cells = end-stopped (true for both complex and simple cells)
segregation between MT and V4 � would/can you bring them together?
if you can�t tell from an object�s contrast/colour where the edge is, motion information might help disambiguate (if the two edges co-incide, would help with segmentation)
binding problem is more problematic when it�s complex � 2 objects in motion (whenever you have >3 neurons at once, need to pair them off, e.g. subject-verb-object)
temporal binding � synchronisation (Singer)
if 2 neurons fire simultaneously, then they�re bound together, by a cell elsewhere being sensitive to the firing of 2 of them
if 2 other neurons were firing together in a different time window, they�d be bound together
rolls = very sceptical of it � very little evidence of synchrony empirically in the infero-temporal cortex in the monkey
best seen in the anaesthetised cat, only then in moving visual stimuli � may be an artifact of the 2 segregated networks using the same information
neurons going to combinations of inputs
form first-order combinations. one neuron which only fires in one combination, and another only fires in another combination (synaptically: competition between neurons and a bit of convergence)
building an implementation with the binding built in to the feature analysers
but would require too many feature combination neurons
helps if you only make low order combinations
analyse it in terms of multiplications of various features
prune the search tree automatically by which occur most often in the natural environment because of synaptic weights (wouldn�t work in a natural environment)
difficult to analyse the statistical content of visual scenes
eigenvector