Transcript Visual system - Ohio University

Computational Intelligence
Perception and attention
Based on a course taught by
Prof. Randall O'Reilly
University of Colorado
Prof. Włodzisław Duch
Uniwersytet Mikołaja Kopernika and
Prof. Oliver University of Connecticut
School of Medicine
Janusz A. Starzyk
EE141
1
Motivation
Perception is comparatively the easiest to understand although for many
specific questions there are no clear answers.
General questions:
 Why does the primary visual cortex react to oriented edges?
 Why does the visual system separate information into the dorsal
stream, connected to motion and representation of object locations, and
the ventral stream, connected to object recognition?
 Why does damage to the parietal cortex lead to spatial orientation and
attention disorders?
 In what way do we recognize objects in different places, orientations,
distances, with different projections of the image onto the retina?
2
EE141
Vision

The hierarchical organization begins in the retina, passes through the
lateral geniculate nucleus (LGN - part of the thalamus), reaching the
primary visual cortex V1, from where it's distributed further.
3
EE141
Visual system
Sight in different types of animals is realized in many ways: a snail has
light-sensitive cells without lenses, insects have a complex eye and
10-30,000 hexagonal facets, mammals have an eye with a retina and a
lens, people have around 120M receptors.
4
EE141
Visual system
5
EE141
Visual system
V1 extends rostrally almost to the lunate sulcus and posterolaterally almost
to the inferior occipital sulcus; the V1/V2 border is met before either sulci.
Pyramidal and
stellate cells
Local axon, double
bouquet, basket,
chandelier, bitufted,
neurogliaform cells
There
are three basic types of neurons in the primate V1 (Fig 12):
Spiny pyramidal cells (excitatory)
Spiny stellate cells (excitatory)
Smooth or sparsely spinous interneurons (almost all are GABAergic).
EE141
6
Visual pathways
Visual pathways:
retina => lateral geniculate nucleus
(LGN) of the thalamus
=> visual radiation
=> area of the primary cortex V1
=> higher levels of the visual system
=> associative and multimodal areas.
V1 cells are organized in ocular dominance columns and orientation
columns, retinoscopic. Simple layer 4 cells react to bands with a specific
slant, contrasting edges, stimulus from one eye. A substantial part of the
central V1 area reacts to signals from fovea, where the density of receptors
7
is the greatest.
EE141
Two streams where?/what?
“where” pathway
orientation
“Where" = large-celled pathway,
heading for the parietal lobe.
direction
depth
Recognized
Object ready
for perception
shape
color
“what” pathway
"What"= small-celled pathway
heading for the temporal lobe (IT).
8
EE141
Two streams where?/what?

Milner and Goodale
(1995): visual
pathways don't so
much determine
where and what, as
much as they enable
action and perception.

There is also the old
limbic pathway,
enabling rapid action
in dangerous
situations (after which
follows a wave of
fear).
Where?
- parietal lobe
What?
- temporal lobe
9
EE141
Two streams
Ungerleider and Mishkin (1982):
there exist two notably divided
pathways for processing visual
information, running from the
eye.
Large-grained PA retina cells, 3
types of photoreceptive cones,
large receptive fields, rapidlyconducting axons, activation for
light in a wide band.
Small-grained PB cells, 1 or 2
types of photoreceptive cones,
small receptive fields, slowly
conducting axons, recognize
color oppositions.
Large-celled pathway: runs to two large-celled LGN layers, it's characterized by a
low spatial resolution, high sensitivity to contrast, rapid signal transfer, without
information about color. The small-celled pathway has 4 small-grained layers in
the LGN, high spatial resolution, color, slower information transfer, low sensitivity
10
to
contrast.
EE141
Retina
 The retina is not a passive camera registering images.
 Crucial rule: enhancing contrasts underlining changes in space and
time, strengthening edges, uniformly lit areas are less important.
 Photoreceptors in rods and cones,
 3-layer network, ganglion cells =>LGN.
Receptive fields: areas, which stimulate a
given cell.
The combination of signals in the retina
gives center-surround receptive fields
(on-center) and vice versa, detects edges.
Each individual field of cells can be
modeled as a Gaussian model, so these
fields are obtained as a difference of
Gaussians (DOG).
11
EE141
Lateral geniculate nucleus
 Signal compression – partly already done in the retina.
 Different types of information find their way to different LGN layers.
 Intermediate station – all sensory signals (except olfactory) go
through different nuclei of the thalamus.
 Dynamic information processing: steering attention and fast largecelled pathway reacting to motion.
 Retroactive projections V1=>LGN are an order of magnitude more
numerous than LGN=>V1 (role - prediction).
The competitive dynamic
selects signals from the visual
field, especially involving
motion.
12
EE141
LGN of the Thalamus
Parvocellular
layers 3-6
Magnocellular
layers 1& 2
13
EE141
Edge detectors
Contrasting signals points from the LGN are organized by the V1
cortex into edge detectors oriented at a specific angle.
Simple V1 cells combine into edge detectors, enabling the
determination of shapes, other cells react to color and texture.
Properties of edge detectors: different orientation;
high frequency = fast changes, narrow bands;
low frequency = gentle changes, wide bands;
14
polarization
=
dark-light
or
vice-versa,
dark-light-dark
or
vice-versa.
EE141
Topography of the cortex
Specific construction, partly genetically determined, partly develops
thanks to stimulation, retinotopic organization like in the LGN.
Different types of edge and texture detectors are topographically packed
in the V1 cortex into hypercolumns, containing separate signals from the
left and right eyes (3D vision, not in all mammals).
Blob region: signals of color and
partly of shape, low frequencies =>
V4.
Interblob region: edge detectors,
every 10o, high frequencies.
Hypercolumn ~1mm2, ocular
dominance columns in layer 4.
15
EE141
Representation in the V1 cortex
Oriented edge detectors can be created by correlational Hebbian learning
based on natural scenes.
What happens with information about color, texture, motion?
16
EE141
Retinoscopic maps in V1
The spatial position of the ganglion cells within the retina is preserved by
the spatial organization of the neurons within the LGN layers. The
posterior LGN contains neurons whose receptive field are near the fovea.
17
EE141
Area V1: The Primary Visual Cortex


V1 is made up of 6
layers (no relation to
6 layers in LGN).
LGN sends axons to
layer IV of V1.
 M and P cells are
separate.
 Right and Left eye
are separate.
18
EE141
Retinal ganglion cells

There are two types of ganglion cells in the retina:
 Large magnocellular ganglion cells, or M cells,
carry information about:
– Movement
– Location
– depth perception.
 Smaller parvocellular ganglion cells, or P cells,
transmit signals that pertain to:
– Colour
– Form
– texture of objects in the visual field.
19
EE141
The geniculostriate pathway




The M cells send their
information to layers 1 & 2.
The P cells send their
information to layers 3-6.
So, layers 3-6 are involved
in processing information
concerning fine detail and
color.
Layers 1 & 2 process
information concerning
movement.
20
EE141
Dorsal pathway
Large-celled pathway:
from the occipital lobe through
the dorsal pathway to the
parietal cortex. Arrives at the
4B layer in V1, from here to the
thick dark stripes of the V2
region, analyzes information
about object motion.
In V1, layer 4B => V5,
localization in the field of vision,
motion.
V5 stimulates the parietal lobe,
PPC (posterior parietal cortex),
regions 7 and 5; this enables
spacial orientation, depth and
motion perception(eye
orientation).
21
EE141
Ventral pathway
Small-celled pathway:
the ventral pathway, to the
inferior temporal cortex.
V1 => V2 interblob region,
reacts to line orientation, gives
a large visual acuity, without
color.
V1 => V3 blob region, reacts to
shapes, reaction to color in the
neurons in the dark stripes of
V3.
V2 => V4, main area of color
analysis, information arrives at
the inferior temporal cortex (IT).
The IT area in the inferior
temporal lobe has neurons
which react to complex objects.
22
EE141
Model v1rf.proj.gz, Chapt. 8
How do receptive fields form? Where do these V1 properties come from?
Inputs: 12x12, signals from LGN
cells: on- and off-center.
Input images: randomly chosen
from a natural 512x512 scene.
Hidden layer 14x14; connections:
coincidental with the input,
excitatory with the surroundings.
Description of the project in Chapt. 8.3.2. Natural shapes and textures
23
lead
to
specific
receptive
fields:
from
this
come
reactions
to
edges.
EE141
Model properties
The V1 cortex receives from the LGN an on/off signal with heightened
contrast, input to V1 through layer 4, processing in this model responds
to overlapping processes mainly in layers 2 and 3.
The model includes one hypercolumn, analyzing a small sector of the
image from images of landscapes and plants => all elements see the
same thing.
Properties: spherical geometry, i.e. top = bottom, left = right;
independent inputs for on/off cells, in accordance with biology;
strong and widespread excitatory horizontal connections – like in SOM;
kWTA leaves ~10% active neurons.
Contrast for weights is small ~1, because these aren't decision-making
neurons, thresholds are large (~2) to force sparse representations,
strong correlations.
Noise helps in avoiding weak solutions.
EE141
24
Exercises from v1rf
Check the structure, connection weights (r.wt): strong activations within
the hidden layer, random connections with on/off inputs.
LoadEnv to load the 512x512 image - for the training 10 images were
used, here is one random one, processed into on/off points.
StepTrain – observe the oscillation of learning for phases – and +
Complementarity of on/off: stronger "on" activation for images which are
brighter in the middle than on the edges, dark = extra "off" activation.
Question: what can we expect if horizontal connections will dominate?
Check your predictions, temporarily changing lat_wt_scale 0.04 => 0.2.
LoadNet to load the trained network, after 100,000 presentations of
images and several days of calculations...
25
EE141
Receptive fields
How do receptive fields form? Where do these V1 properties come from?
Check the activation r.wt, change the color
scale so we can better see the field
orientation, check several hidden
elements, bi- and tri-polar fields of both
types.
Load all: View, RFIELDS
activation on=red, off=blue.
Orientation, position, size, polarity are 4
different traits of receptive fields.
Radial orientation changes (pinwheels),
singular points.
View, PROBE_ENV shows 4 different probe stimuli, StepProbe will
26
show
activation
of
hidden
units.
EE141