Lecture 4- Representation of Complex Images

Question 1

How does information from secondary somatosensory cortex differ in the dorsal and ventral streams?

Answer 1

• Information from the secondary somatosensory cortex contributes to the dorsal stream by specifying the movement needed for grasping a target
• Information from the secondary somatosensory cortex contributes to the ventral stream by providing information about object size, shape, and texture

Question 2

What areas of the cortex are associated with the ventral stream?

Answer 2

Goes from the primary visual (striate) cortex across the prestriate cortex into the inferotemporal cortex (IT cortex)

Question 3

What do neurons in the monkey IT cortex respond to?

Answer 3

To complex objects like faces and fire maximally when this face is at a certain orientation (side on)

Question 4

How was a ‘Jennifer Anniston’ cell found in the human IT cortex and what does this mean?

Answer 4

• Showed 100 different objects and found a cell that fired a lot when presented a picture of Jennifer Anniston
• It didn’t respond to famous people in general or to just a human face only Jennifer Anniston

Question 5

What was found with the ‘Halle Berry’ cell that expanded what was found with the ‘Jennifer Anniston’ cell?

Answer 5

• Cells responded not only to pictures of Halle Berry but to a mask of cat women and also her name. This posits the idea that there is a cell for the concept of Halle Berry not just how she looks
• Of the different modes the Halle Berry was presented the name elicited the most response, then photos, then cat women, then drawings of Halle Berry

Question 6

What does the ‘Steve Carell’ cell show?

Answer 6

It’s not just female faces that result in specific cells it’s men to

Question 7

What is invariance? What different types of invariance are there?

Answer 7

• One striking aspect of the Quiroga results is the consistency of responses across different images of the same person or object. This is known as ‘invariance’ - Encoding a representation so that it is identified regardless of size, orientation, colour etc. The cell fires/ responds the same either way.
• Types of invariance: translation, rotation/ viewpoint, size, illumination

Question 8

What does lots of invariance suggest?

Answer 8

It’s a high order cell not just simply responding to a component like shaped but a concept.

Question 9

Do we have specific cells for things other than faces?

Answer 9

Yes, there is an architecture discriminating cell

Question 10

Could cell firing be modulated by familiarity?

Answer 10

Yes: categories were self & family, experimenters, famous people, non-famous people

Self & family + experimenters was high compared to famous and non famous people. Within the later two categories famous was significantly higher than non famous.

Question 11

What was found in relation to the experimenters?

Answer 11

A neuron in the amygdala with a selective excitatory response to an image of one of the researchers running studies with the patient at UCLA.

Question 12

What is another theory as to well cells are firing for certain objects/ people? How was this tested?

Answer 12

• Could simply be an emotional response
• Put images of things people typically find scary like spiders and snakes into image database. These did not ellict as much firing though as the faces.

Question 13

What was interesting about the ‘Jennifer Anniston’ cell?

Answer 13

It didn’t fire when she was shown with Brad Pitt

Question 14

What are grandmother cells? What type of coding do they utilize?

Answer 14

• 1 small group of cells firing when see a certain thing
• This firing in these specific neurons gives the perception of that thing
• Local coding= Self-contained: only requires a small number of cells

Question 15

What is good about local coding theory? What are the problems?

Answer 15

-Good at discriminating between objects

Problems:
1) requires huge number of gnostic units (you need 1 grandmother cell for everything you can perceive)
2) susceptible to damage (if the cells get damaged you lose the ability to identify that thing, damage doesn’t tend to occur this specifically)
3) how to you perceive novel objects? (would need a bunch of blank cells sitting there)
4) In the Quiroga et al study what is the probability of finding the
one ‘grandmother’ (Aniston) cell out of the millions* available
with just 100 or so images? (suggests there are actually many Jennifer Anniston cells and we just happened to find one)
5) Generalization is difficult (if you were to have separate pockets of information in grandmother cells)
6) Pattern completion requires access to representations of other
similar objects (not separate units as such)

Question 16

What are the critical goals of visual information processing?

Answer 16

To separate patterns, to complete them and to allow for generalization: important for survival. E.g. recognizing a tiger in the grass/ or partially hidden/ or in a weird shape. You still need to be able to recognize it’s a tiger no matter how distorted or occluded it is.

Question 17

What is pareidolia and what does it show?

Answer 17

• Seeing faces in inanimate objects

- Shows the brain interpreting info and completing patterns in meaningful ways

Question 18

What is an alternative to the grandmother cell/ local coding model?

Answer 18

• Dense (ensemble, population or distributed) encoding theory
• There is no one cell that is representing your grandmother but cells responding to different components of her face
• The activity is distributed across a whole group of neurons
• The collective pattern of activity across the cells corresponds to the image/ identification of your grandmother

(NOTE: just using the example of your grandmother as the image/ thing to be identified, this is not related to grandmother cells)

Question 19

What does dense coding allow for?

Answer 19

Generalization
e.g. between Steve Carell and Jennifer Anniston a lot of the cells responding are the same (they are both faces and have certain features) only few activated cells are different but this is what creates the ability to distinguish between the two (the overall pattern of activity is altered)

Question 20

What trade off occurs in the representation of complex images?

Answer 20

-Pattern Separation, and Pattern Completion and Generalisation.
The nervous system has to balance out these priorities of the visual system. It does this via using some dense encoding (allows for pattern completion and generalization) and some local coding (pattern separation).

-Combined use of both local coding and dense coding creates an intermediate coding scheme where representations are distributed but maintain some degree of sparseness.