Please enable JavaScript.
Coggle requires JavaScript to display documents.
MODULAR AND DISTRIBUTED VISUAL REPRESENTATIONS :PENCIL2: B+B LECTURE 2 -…
MODULAR AND DISTRIBUTED VISUAL REPRESENTATIONS
:PENCIL2: B+B LECTURE 2
Modularity and the mind and brain
History
Phrenology (Gall, 1835) - location of different capacities, moral, aspiration, etc. Lead to different deformations of the skull.
Principle Idea
Different capacities of the mind are organised into different 'modules' (Fodor, 1983).
Compartmentalising the mind into different functions.
Modules are domain specific (only operate on specific information)
Modules are functionally encapsulated (their computations are independent of other modules)
On a neural level
Modular organisation should be evident in distinct neural populations specialised for particular functions.
Assume that, for example, visual functions (object recognition, face recognition, scene perception) are all organised in specialised modules.
Mapping
Facial and body processing = lateralised to the RH
Word processing = lateralised to the LH
Object-selective cortical regions
The aim is to identify brain regions that preferentially respond to meaningful objects.
General approach = record fMRI and compare objects to a set of control stimuli (e.g. scrambled objects with similar contrast, brightness etc. but different configuration)
Basic idea
Scrambled objects activate low-level visual cortex (e.g. V1) in a similar way, but object-selective regions should more strongly be activated by the intact objects, if such regions exist.
Object-selective regions can therefore be identified via subtraction.
Activation in the early visual cortex should be equal, because we have controlled for things like contrast and luminance, and therefore cancel out these factors in subtraction, only leaving the object-selective cortex.
Regions preferentially responding to objects are found in the lateral occipital cortex (LO) in both hemispheres.
Found in similar regions across people.
Category-selective modules
Aim to identify modules that preferentially respond to different categories of objects (e.g. faces)
General approach: Record fMRI and compare objects of the chosen category to objects of other categories.
Why? Faces need to be analysed very thoroughly to have successful social interactions.
Contrast faces with non-face objects.
FINDINGS = A face-selective module is consistently found in right fusiform cortex; Fusiform Face Area (FFA)
Fusiform cortex extends along the ventral temporal cortex, and is more strongly activated in the RH when faces are shown
For most categories, there are more posterior versions (e.g. the occipital face area and the extra-striate body area), and anterior ones (e.g. the fusiform face are and body areas)
EBA and OFA do more localised perceptual analysis of faces such as analysing the eyes alone, or analysing how the eyes are configured.
Whereas the FFA potentially does analyses that are more holistically processed.
Posterior = localised
Anterior = holistic
How important are these modules for perception?
Pitcher et al (2009)
Disrupted OFA, LO, EBA using TMS during a discrimination task involving two similar looking objects or faces or bodies.
Detrimental effect on perception of objects within that category.
Computations in category-selective regions
FFA activations show graded sensitivity to real faces, symbolic faces, face parts and other objects.
The more facial features that are visible, the higher the response.
Category-defining features
Reduce stimuli to essential combinations of visual features to understand the tuning of category-selective regions.
Caldara & Seghier (2009)
OFA is activated more strongly when 'tennis racket' stimuli are composed in a more face-like manner (e.g. symmetrical, dense features at the top)
Representations of perceived category
Category-selective regions are more active when the same physical stimuli is perceived as belonging to the regions preferred category, suggesting that it’s not only about visual features, but about category membership.
For example, “Mooney” faces activate the FFA, but not PPA, more strongly when they are recognised as faces than when they are not
(Andrews & Schluppeck, 2004).
Parahippocampal Place Area
Located in the parahippocampal gyrus
Typically found if you compare activation of houses or scenes to scrambled versions, or if you compare them to other objects/faces.
Basic idea:
Scenes are more rectilinear (i.e. more right angles) than other stimulus classes (such as faces or bodies)
Therefore, PPA perhaps just is a ‘rectilinearity detector’ that is tuned to the presence of right angles, assembled from orientations coded in early visual cortex.
Rectilinearity index = how many right-angles the stimulus has.
One hypothesis is that PPA detects this rectilinearity in the stimuli, and therefore is more active for scenes.
You could imagine PPA being connected to an early visual area, for example, the primary visual cortex.
PPA then combines these simple orientation features into more complex ones.
E.g. combining a vertical orientation with a horizontal orientation into a conjunction of the two (tuned to a right-angle)
To test this idea, the researchers showed participants stimuli that were more or less rectilinear.
PPA consistently preferred rectilinear over round-ish stimuli, suggesting that PPA indeed preferentially responds to rectilinear stimulus features.
Not about scenes anymore, it’s about probing these features specifically.
A reductionist explanation for the existence of PPA.
Visual feature anaysis
PPA Supports Spatial Cognition
Basic idea:
PPA is responsible for efficiently representing the space around us for enabling us to act upon it.
Should be tailored around tasks such as navigation.
If this is true, PPA should not primarily be driven by visual features, but by the presence of a visual scene that affords spatial cognition.
To test this, researchers showed participants simple objects or scenes assembled from building blocks.
Don't need visual input to activate PPA.
Sheds doubt on visual feature hypothesis.
However it is possible and probable that people are visualising the scene that they are touching.
Solution: Test on the blind.
Repeat the experiment with blind people and let them touch the building-block scenes.
Even in the blind participants, you get a stronger activation when they touch the scenes as opposed to the objects.
Pretty strong evidence that these regions are not actually coding for the presence of particular visual features, but maybe the evaluation of scene contents?
Distributed Representations
As an alternative to the modular organization of visual cortex, it has been proposed that objects are coded in a distributed fashion, where the response patterns across many neurons (and regions) are diagnostic of what we see.
Under this view, modules may only be clusters of local selectivity (that show up akin to the tip of an iceberg).
Every stimulus would only be represented in distributed activity across a number of nodes that do not need to be spatially close to each other.
To test this idea, lets show fMRI participants images of different categories again (e.g. faces and houses)
Let’s do this across multiple runs of the experiments and then see how well response patterns (across fMRI voxels) across the whole ventral visual cortex discriminate between the faces and houses (Haxby et al., 2001)
One way of doing this is to split our data into two chunks, for example the odd runs and the even runs.
Then let see whether response patterns to faces in the odd runs are more similar to faces in the even runs than to houses in the even runs.
If this is the case, faces and houses can be discriminated based on these large-scale response patterns.
Take-home message
Distributed activity patterns contain category information and the distributed code seems to be important for object representation in the brain.
Tracking representations across time
We can perform the same analysis using MEG, exploiting the method’s temporal resolution.
At each time point, we can record response patterns across all MEG sensors every millisecond.
0ms, information hasn’t arrived to the brain yet.
We can again correlate the response patterns for each pair of objects, yielding a measure of similarity across time.
We can further deconstruct these similarity matrices, tracking when specific categories become discriminable.
These analyses reveal that basic level category distinctions emerge before more fine-grained ones.
This suggests a progression from visual to conceptual category representations.
Is the brain really modular?
There seems to be a co-existence of category-selective modules and distributed codes that carry information about visual objects.
These different codes may reflect complimentary representations that are used for different types of real-world tasks (e.g. understanding the shape of an inanimate object versus individuating familiar people).