B&B - Vision II
Infero-temporal cortex and visual agnosias
weiskrantz + saunders
can learn object discriminations without your IT
but monkeys are slower
couldn�t do the transforms + invariance
supported by neurophysiology
holmes and gross
prefer vertical pen to horzontal pen � rely only on whether it�s vertical?
we don�t know to what extent higher primates �see� things as being anything
new reference
paper in Neuron � by Rolls � to be published in the next year
re neurophysiology of invariance
no translation invriance in V1, but you do in IT � done it for faces, objects
1998 Rolls & Booth, incl view-invariant
10 objects in the monkey�s cage
some Ns fire for the same object in different views
no longer think of things as being 3D object-centred, because theory suggests 2D views
lumination invariant � few experiments done so far
colour � respond similarly in grey scale
small variations in shape �
Tanaka � perspective invariance, not for real objects, altered perspectives of outline shapes, showing that the IT neurons still fired whatever shape you use
occlusion + completeness � often respond as the same object
texture � spatial frequency
fourier analysis of face to eliminate low frequencies � would like like a pencil drawing, so you�d lose all the bits where there�s no change, e.g. surfaces
opposite � blurred � cells also respond, show invariances even if thre�s no overlap = spatial frequency invariance
size � mean size tuning (12.5 times size, and still carry information about the faces according to the half-amplitude statistical measure)
= solving a massive computational problem in 4 stages
in <16fps, = approx 90ms to get to the IT, = 50ms of cortical processing time
= about 15ms per stage
neural network � integrate and fire � 17ms
Visnet demonstrates the computability in a 4-layer network � doesn�t store enough variations of every size/shape
Wallace & Rolls on Visnet (1997?), new journal in Neural Computation
we know how the information is represented in it
can read off the activity of a handful of cells, relate back to stimulus
by knowing what the firing rate is, can guess from the average firing rate which of 8 possible stimuli it was
= obviously different for every animal
every new neuron linearly adds to the information, so every N carries independent information � can be decoded by doing the neuronal dot product (Rolls et al 1997, NN&BF)
doesn�t matter which cells you take � works for random selection
done with faces + objects
10 objects from 4 views in the monkey�s cage
dot product of firing rate of cells with the average firing to every stimulus ever tested, can tell which stimulus it was
don�t take into account the temporal aspects
which are crucial at this level for reading it
I can�t do that for a new brain � is this a pseudo-problem?
IT � orbitofrontal, amygdala, BG, so they understand the readout
well, how does the orbitofrontal cortex read it out � pattern associator learns it and can read the code, and can do a visual/taste co-ordination
what else do you need? have we understood how the code can be used? need we posit a homonculus?
how does that work for complex composite objects?
what if you have a visual scene? do different objects interfere with each other
single object ID in parallel, serial ID of objects in scene
if 2 objects: weight towards the fovea
translation invariance will � in a real scene � if it is really translation invariant in the IT, how do you know which is the target for your action
background invariance
does the receptive field get smaller?
why is this code invariant?
why not have one size/position/etc. coded by one set of neurons?
would need to bring them together to compare object at different sizes � pattern associator wouldn�t recognise the same object in different positions and that wouldn�t be adaptive
need to compute the object before sending it off to an area which decides what to do with it
processing/storage requirements
we know a bit from lesion and single-cell studies
apperceptive � can�t copy or match
do you have enough information coming into your spatial discriminator?
is it a reception or a form analysis problem?
associative
discrimination of objects = good, can copy + match
can�t access the semantic/naming level
if semantics without naming = anomia (need to show that they do first have the concept to distinguish it from a general semantic loss)
do associative agnosics have an IT?
is it the output from that that is impaired?
compare it with functioning IT-orbitofrontal transmission
weiskrantz study � can they do that?
whether they could match/copy different sizes/views of objects � say whether or not they�re the same?
if they can, it�s a disconnection at a higher level
this hasn�t been done
IT in humans = fusiform face area
low
normal visual acuity
reception
form analysis
high
simultanagnosia � usually higher level damage to parietal, re < visual STM (but not auditory STM problem)
has there been a Brain & Emotion review?
next week � Dorsal (parietal)
what are the functions of the parietal cortex?
Kandel � Principles (esp Richard Anderson)
get rolls to email us
rewrite this essay