INF1-CG VMA Lecture 2: Neurons (for vision): photoreceptors, ganglia, etc.

1. Machine representation and implementation: Why is machine vision hard?

Examples of machine vision:

• automated inspection systems in factories
• face matching/face recognition
• extracting shapes and planes from images
• Identifying (exactly, or as to category) objects

Despite more than half a century of effort

• most machine vision applications are poor compared to their human equivalents
• Raw image data is often very noisy
  • or of poor quality (especially for 3D data)
• There is often too much data
  • so that computation becomes overly difficult or time-consuming

That's about implementation

But we're losing even at the representation/algorithm level as well

• Object/pattern recognition
• figure-ground segmentation
• salient (important) region detection
• visually guiding robot/vehicle navigation or other actions

2. The Challenge of Vision

Some questions which will be addressed in the following lectures on vision:

• How does light in the world get into your eye and become information/ representation in the brain?
• How do we extract useful information from this representation?
• How are the needs of an organism reflected in its visual system?
• Why could you starve your pet frog even if you put food on the ground all around it?
• Our visual systems often get fooled. what can illusions tell us about visual processing?
• 
  Public domain
• And last but not least, "Why is there a neuron in my brain that only responds to pictures of Halle Berry?" (weird but allegedly true)
• 
  Source unknown

3. Visual processing roadmap

Visual information starts as very small units

• Single cells in the retina detecting light
• Light information gets passed through layers of the brain
  • to become richer pieces of information
  • Say, the presence of a line
  • ...and even richer pieces (this is an edge of something)
  • ...all the way up to object recognition (this is a chair)

Not the whole truth. . .

• This course presents a brief and simplified version of neurons
  • and the early visual processing system
• in order to introduce our approach to cognitive modelling
• We won't discuss color vision or depth perception at all

4. Neurons

Neurons are specialised cells

• connected with other neurons
• which rapidly transmit information

Neurons have a specialised cell body

• with many bushy dendrites
  • which take incoming connections from other neurons
• and one axon (or nerve fiber)
  • which makes outgoing connections to other neurons

Axons and dendrites carry electrical signals called action potentials

• Which are brief spikes of voltage

We say a neuron fires

• when it raises an action potential on its axon

An axon-dentrite connection can be excitatory

• tending to make the target neuron fire too

or it can be inhibitory

• tending to make the target neuron not fire

5. Photoreceptors and visible light

A receptor is any neuron specialised to respond to energy from the environment (light, pressure, molecules)

The first layer of neurons in the retina of the eye are called photoreceptors

• Photoreceptors fire in response to detecting a required (threshold) amount of visible light
• Many wavelengths of light (UV light, X-rays, infrared) cannot be seen by humans
  • Other animals may see more, or less
• We see only a small part of the whole electromagnetic spectrum

Other retinal neurons do not respond directly to light

• But to excitatory and inhibitory messages from the photoreceptors

6. Convergence: Passing messages between layers

Neurons in the vision-processing areas of the brain (the visual cortex) are organized into layers

• Every neuron in a higher layer receives messages from many neighboring neurons in the layer below
  • These higher neurons will fire if they receive sufficient excitation
    • without too much inhibition
• This phenomenon is known as convergence
  • Neural convergence is essentially a method of compressing and simplifying visual information.

7. It's not all one way

• Neurons not only connect directly to the next layer up
  • but also connect laterally to neurons in the same layer
• Neurons from later layers in some parts of the brain also have back projections that pass information to lower layers
  • But we won't cover those in this course

Both upward and lateral connections can be inhibitory as well as excitatory

• You can think of neurons as summing their inputs, positive (excitatory) and negative (inhibitory)
• And firing if the result exceeds some threshhold

8. Receptive fields

Some higher-layer neurons fire based on activity in their receptive fields

• That is, areas of the retina which contain multiple photoreceptors
• All connected to the higher-layer neuron in question
• Think of this as the area the neuron can ‘see’

Such a neuron responds to patterns of light, not just its presence/absence

• Different types of neurons respond to different patterns

Generally, the higher up in the processing stream

• the more complex the pattern

9. Receptive field example

A ganglion is a higher-layer retinal neuron

• with a circular receptive field

• The illustration here shows fields that respond to a spot of light
  • surrounded by darkness (center surround cells).
  • +   Photoreceptor with excitatory connection to ganglion
  • -   Photoreceptor with inhibitory connection to ganglion

10. Serendipitous science: Hubel and Wiesel

Early research had discovered the center-surround receptive fields of ganglion cells

• sensitive to spots of light

Hubel and Wiesel were interested in extending this type of research

• In the late 1950s

They focused on recording from individual neurons

• in the visual cortex of cats
• This area is also known as the striate cortex or V1

Cats have receptive fields fairly similar to those in human visual processing

H&W found no response to any of the dot-like stimuli they tried

• but they did find cells fired as the edge of the slide was being inserted into the projector
  • In other words, a line detector

11. Hubel and Wiesel, cont'd

The true importance of this research was that it demonstrated that different neuron types could be viewed as steps in a processing hierarchy.

• Validating the idea that there are multiple steps between what is in the world and our internal representation

For example, some cortical cells responded to spots anywhere in a narrow rectangular region

• in a particular orientation

They also responded, more strongly, to a rectangular stimulus

• Covering the same target area
• Or to a group of dots arranged at the same orientation
• Ganglion cells were known to respond to spots of light.
• So we have evidence for three layers now

They also found fourth-layer cells with complex receptive fields

• For example, summing together multiple edge-detectors across a region of the retina

Hubel and Wiesel's work on orientation selectivity (along with collaborator Roger Sperry) won the Nobel Prize for medicine/physiology in 1981

12. Columnar organization and tuning curves

Neuron which respond to lines and edges

• Will usually fire strongly for lines of a particular orientation
• Will also fire less strongly for lines with a similar (but not ideal) orientation
• Won't fire at all for stimuli with very dissimilar orientations.

This pattern of responding is known as a tuning curve

• It also applies to many stimuli more complex than lines
• As well as the simpler case of spot detection we saw already

Neurons directly on top of one another in the cortex (members of the same vertical column)

• tend to respond to stimuli with the same orientation

Adjacent columns have similar preferred orientations.

13. Admin

No change of venue for tomorrow

• We're still in AT 2.12
• On the 2nd floor of Appleton Tower

Course texts (required, for details see this week's reading list):

• Sensation and Perception (Goldstein 2007 or 2010)
• Vision (Marr, W.H. Freeman & Co, 1982. Also reprinted 2010)

Original papers (optional reading):

• “Receptive fields of single neurons in the cat’s striate cortex” (Hubel & Wiesel, 1959)

For more detailed information on neurons and neurobiology:

• Neurobiology (Shepherd, 1994.) There are probably more recent editions of this
• Scholarpedia entry for “Neuron” at http://www.scholarpedia.org/article/Neuron