ML3

Question 1

Why study early vision? Name 3 reasons

Answer 1

1. Image processing is a common application of artificial deep networks
   (particularly object recognition)
2. Early vision is very thoroughly investigated
   (Visual input is easy to control)
   (Good animal models of human vision)
   (Computational aspects are well understood)
3. Common application of DCNN as simulations of neural processing

Question 2

We believe that very similar principles operate in the networks in the brain like the retina. Why?

Answer 2

We already saw how the retina separates images into multiple feature maps for different colours and spatial frequencies.

Question 3

Organisation of early visual cortex. Explain the process

Answer 3

The information from the two eyes is joined together. The left half of both retinas goes to V1 of the left hemisphere, and the other half to the right hemisphere.

Question 4

Do spatial relationships from the image become spatial relationships in the cortex?

Answer 4

Yes

Question 5

Why is it important that relationships from the image become spatial relationships in the cortex?

Answer 5

• This allows filtering and integration by dendritic trees with a limited cortical extent, a limited spatial extent in the feature maps.
• It also allows interaction over minimal distances: cells that wire together lie together for efficient interactions.

Question 6

Name the 3 different cell groups carrying information from the retina

Answer 6

parasol, midget and bistratified cells

Question 7

Are the parasol, midget and bistratified cells helt in slightly different cell populations at different depths or layers in the grey matter?

Answer 7

Yes

Question 8

Are grey matter layers equivalent to neural network layers?

Answer 8

No, but here we do see several neural network layer - like stages, i.e. different numbers of synapse steps from the photoreceptor input, are sometimes in different layers of grey matter in the same cortical area.

Question 9

Do the spatial relationships allow these neurons to interact most easily when they are carrying information on the same feature type, as these are closest together?

Answer 9

yes

Question 10

Where do responses to specific edge orientations first emerge?

Answer 10

In V1. These cells respond when an edge has a specific orientation (the preferred orientation) and is shown in a specific position.

Question 11

Are the filters operating on image inputs?

Answer 11

No, the inputs are already outputs of the previous layer of filters

Question 12

Why is it computationally useful that neurones that have similar responses are found together?

Answer 12

If we have a filter that will accept a range of orientations, the neuron implementing this filter can synapse with a group of nearby neurons.

Question 13

Orientation selectivity

Where is contrast initially computed in an orientation-independent filter?

Answer 13

In the retinal ganglion cell

Artificial DCNNs often avoid this step, going directly from image to orientation

Question 14

Where are the orientation-selective responses computed?

Answer 14

V1, by operations comparing these RGC outputs (not image content)

Question 15

What happens with the orientation preferences?

Answer 15

They gradually change across the cortex, at a much finer scale than the spatial visual field maps

Question 16

Orientation preferences gradually change across the cortex. This squeezes multiple feature maps into the same 2D cortical surface. Why?

Answer 16

This may arise to optimise connection lengths.

Only when filters are not shared across the network layer’s image representation

Question 17

Why is it rare for artificial deep networks to consider changes over time?

Answer 17

Because they are generally given single images as inputs rather than movies, largely because of computational limitations.

Question 18

Do biological visual systems inherently receive inputs that change over time? What do they need?

Answer 18

Yes, they have a mechanism to detect motion, its direction and its speed

Question 19

What can a filter have (regarding motion detection in primary visual cortex)?

Answer 19

time dimension, and spatial dimension

Question 20

Name the different responses that we possess?

Answer 20

Colour, orientation and motion in different neural populations. We normally perceive all of these together

Question 21

Primary visual cortex also receives separate inputs from the two eyes, Dublin g up the feature maps. True or false?

Answer 21

True

Question 22

Binocular integration

Answer 22

(2 eyes). This underlies a large part of our depth perception.