Cogs 160: Neural Coding in Sensory Systems
Cogs 272: Computational Models of Sensory Coding
Prof. Angela Yu
Department of Cognitive Science
April 1, 2010
http: www.cogsci.ucsd.edu/~ajyu/Teaching/Cogs160_sp10/cogs160_272.html
Grading
Participation (reading, discussion): 25%
Mini thought essays (1 page, weekly): 75%
Extra credit: respond to a bonus question
The Eye and the Retina
Light is focused by the cornea and the lens onto the retina, a thin
layer of neural tissue at the back of the eye which contains
photoreceptors. Photoreceptors transduce light into neural signals
and pass their signals on to the brain.
Anatomy of the Retina
Retinal Ganglion Cells
Left cell is ON-center/OFF-surround.
Right cell is OFF-center/ON-surround.
Neurophysiology of On/Off Pathways
light
Neurophysiology of On/Off Pathways
This View is Far Too Simplistic!
Left cell is ON-center/OFF-surround.
Right cell is OFF-center/ON-surround.
THE ROLE OF THE RETINA
IN VISION
Jake Olson
WHAT DOES THE RETINA DO?
• Popular
• But
opinion: Filter
does it also compute?
Anatomy of the Retina
Outer Plexiform Layer
Inner Plexiform Layer
Retinal Ganglion Cells
Left cell is ON-center/OFF-surround.
Right cell is OFF-center/ON-surround.
SO, IS IT JUST A FILTER?
•
Just passes the visual image to the cortex for the cortex
to processes it.
•
Simple spatiotemporal filtering.
•
Adapts to light
•
Center surround inhibition clarifies location.
•
But then why 50 types of cells?
•
So maybe the retina does more?
COMPUTATION!
Perhaps each ganglion cell already computes a specific
visual feature?
• If so, a neural circuit would be needed to make the
calculation, one for each ganglion cell type/different
calculation.
• Instead of sending a filtered “pixelated” version of the
visual scene to V1, retina sends extracted features.
• Benefits:
• enables speedier processing.
• explains the multiple cell types.
• direct projections for quick actions are possible.
•
DO WE HAVE ANY EXAMPLES?
• Well known: direction-selective ganglion cell.
• Light detection
• Motion Detection
• Texture Motion - “Y” type
• Object Motion
• Approaching Motion
• Anticipation
• Motion extrapolation
• Omitted Stimulus
• Saccadic Vision
• Saccadic Suppression
• Latency Coding
• Adaptive Computation
DIRECTION SPECIFIC
GANGLION CELLS
• Respond
primarily to movement direction of stimuli
in their field.
• Direction
specific - only fire when motion is in a
certain direction.
Taylor & Vaney
WHY IS THIS EXAMPLE
IMPORTANT?
• Computes
•A
a specific feature.
large amount of stimulus information is discarded.
• Computation
• No “higher
is explicitly available in the output of the cell.
processing” needed.
LIGHT DETECTION
•
Rods can detect one photon, humans sense a flash of light of just a few
photons.
•
Ganglion Cell gathers from hundreds of rods.
•
Noise is present in the circuitry.
•
How does retina separate the signal from the noise?
•
•
Temporal filter.
•
Threshold needed to be summed.
Specific to rod-bipolar pathway.
MOTION
• What
kind of motion is experienced on the retina?
• global
motion - noise
• object
motion - signal
“Y-TYPE” GANGLION CELLS
• Fire
when a texture moves.
• Largely
layout.
independent of motion direction or spacial
OBJECT MOTION SENSITIVE
GANGLION CELLS
•
Constant “noise” movement.
•
Need to compare local motion to overall motion.
•
Center Surround receptive field.
•
Projects to Superior Colliculus in mice.
APPROACHING MOTION
• Off-type
• Works
Response - Driven by increasing dark area.
despite global brightening of the scene.
ANTICIPATION
• Very
useful.
• Experience
the past.
• Phototransduction
delay = 10s of ms.
MOTION EXTRAPOLATION
Motion is typically in a smooth trajectory.
• Experiments show ganglion cell activity actually is at the
true location of the object.
• This is due to spatiotemporal characteristics of the
ganglion receptive fields and the dynamic gain-control
mechanism.
• Crude estimate, does not account for speed.
• Gain control depends on stimulus contrast.
• Problems?
• Turns, Reversals.
•
OMITTED STIMULUS RESPONSE
• Expect
a repetitive stimulus.
• Temporal
anticipation instead of spatial.
• Multiple
conflicting theories, no definitive evidence in a certain
direction.
Blue - Response
Red - Stimulus
SACCADIC VISION
•
Short periods of fixation, then
sudden, rapid movement.
•
Global image movement for
ten’s of ms.
•
Saccadic suppression
•
Could it use global vision
information?
LATENCY CODING
•
Humans can recognize image content in the first 100 ms. How is it so fast?
•
First spikes could be informative.
•
If the photoreceptors turn darker, the ganglion cells respond sooner.
•
Sends basic shading very quickly.
SWITCHING CIRCUITS
•
Powerful computational device, allows dynamic routing of information.
•
Can one ganglion cell type do more than one thing?
•
Inhibit one input pathway and excites the other.
ADAPTIVE COMPUTATION
• Retina
receives minimal input from the brain. What
would drive changes in how features are computed?
• For
example, light adaptation.
CONTRAST ADAPTATION
•
Higher contrast = lower sensitivity, higher kinetics.
•
Fast response = .1 ms - “Contrast Gain Control”
•
•
Sensitivity drops some, kinetics up a lot.
Slow response = many seconds - “Contrast Adaptation”
•
Sensitivity drops, kinetics very little change.
PATTERN ADAPTATION
• Slow
adaptation.
HOW IS THIS HAPPENING?
• What
changes?
• Some
bipolar, some amacrine, big effects in ganglion cells.
WHY?
• Fast
- Local statistics
• Slow
- global
FUTURE?
• Reverse
engineering is hard!
• Fluorescent
cells!
DISCUSSION!
•
Messing up the image! How can brain get the whole image back?
•
Does this apply to humans too? Not just simple animals?
•
Center surround issue.
•
Why call this computation and not what the retina was doing before?
•
What are concerns when choosing stimuli for an experiment such as those described?