Transcript Master Thesis Presentation.

Neural Visuomotor Controller for
a Simulated Salamander Robot
Biljana Petreska
Diploma Thesis – March 2004
Responsible
Prof. Auke Jan Ijspeert
Goals of the Project
 Investigate through simulations tightly coupled with neurobiological
data, the neural mechanisms underlying visually guided behaviour in
amphibians
 Implement a closed-loop with the environment onto the existing
neuromechanical simulation developed by Ijspeert, by adding
biologically inspired models for parts of the salamander brain
 Develop a controller that accounts for observations in feeding
behaviour, including prey localization and prey recognition
 Study a model proposed by Ijspeert of structured mapping between the
optic tectum (primary visual processing center) and the brain stem
(motor centers) as a solution to the visuomotor coordination
Interests
 Relevant for perceptual robotics

decoding the brain processes, assigning meaning to complex patterns
of sensor stimuli may lead to the solution of many robotics tasks
 Test bed for probing neurobiological contributions

ideal for the validation or refutation of new theories
Overview
 Short introduction on relevant topics and previous works
 Implemented Models
 Respective Results
 Conclusion and Future Work
Everything you’ve always
wanted to know on Salamanders
 Amphibians
 Great variety of species (3924
indexed so far), sizes (from
16mm to 1.5m), aspects and
lifestyles (terrestrial and/or
aquatic).
 A relatively simple neural
circuitry that presents all main
vertebrate features
 Tractable from an experimental
point of view: an important
amount of behavioral, biological
and neurological data exists
Visually Guided Behavior
 Vision is by far the most important feeding guiding sense. Under good
visual conditions the other signals such as olfactory are overridden
 Feeding strategies (some species can switch from one to another):
 “hunter” strategy: active search for prey. Prerequisites are a short
massive tongue and poor visual capacities.
 “ambush” strategy: wait until prey comes close. Prerequisites are a
highly specialized projectile tongue (up to 80% body length),
evolved visual system and frontally oriented eyes
Visually Guided Behaviour
 Sequence of feeding behavior





orienting
approach
olfaction tests
gaze stabilization
snapping
 Prey preferences (in order of importance)






stimulus size
stimulus velocity
stimulus-background contrast
stimulus shape
movement pattern
experience-dependant
Morphology of the Salamander
Brain
 Functional differentiation of the
brain: structurally different regions
accomplish different tasks
 Global top to bottom visual
information processing
 Principal components:
photoreceptors, retina, optic tectum,
nucleus isthmi, pretectum,thalamus,
medulla oblongata and brain stem
Retinal Ganglion Cells
 First layer of visual processing, transfers visual signals to the brain via
the optic nerve
 3 Types of retinal ganglion cells that project to particular layers in the
optic tectum
Optic Tectum
 Main visual processing center. Integrates also multimodal perception,






such as ascending somatosensory, auditory, olfaction and vestibular
Stratification in 9 layers, first three are retinal afferents
Six morphological neuron types identified (one interneuron)
Topographic representation of the visual field
Viewed as a set of partial overlapping maps due to the different types of
tectal projection neurons
Number of tectal cells in Hydromantes Italicus : 92 000 and 3300 out of
5000 projection neurons are descending
Projection patterns: from and to the retina, pretectum, thalamus,
nucleus isthmi and medulla (reaching the spinal cord)
Distribution and Receptive Field Sizes of Tectal Neurons in H.Italicus
Pretectum
 Has been ascribed a role in optokinetic nystagmus, figure-background
discrimination, pupillary reflex, fixation, phototaxis, and prey-enemy
distinction.
 Properties of pretectal neurons:
 Homogenous arborisation
(no classification was possible)
 Divergent projections
(including to the tectum and spinal cord)
 Large receptive fields
 Receive direct and indirect (from tectum) retinal input
 Direction-sensitive neurons (predominantly in temporonasal
direction)
 Respond to stimuli in the contralateral visual field
Lesion Experiments
Give insight of the function of the destroyed brain region:
 Lesion of the optic tectum: both visual prey-catching and predator
avoidance fail to occur. Local lesions produce scotoma, total blindness
for a part of the visual field corresponding to the size of the lesion.
 Lesion of the pretectum: locally facilitates feeding and abolishes preypredator discrimination, attack everything that moves including their
own extremities and threatening stimuli
 Lesion of the thalamus: unable to avoid collision to a vertically stripped
barrier, affects the binocular field
 Lesion of the medulla oblongata: affects distance, elevation or
horizontal eccentricity estimates, overshoots prey or snaps only in
frontal positions => different components of the stimulus position are
handled through different pathways
Difficulty: in some cases the animal recovers shortly after the lesion and
the relative precision of lesions may induce errors
Previous Works
 Based upon the principle of coarse coding (Eurich et al, 1997):




Motivation: the high sensory resolution observed in nature seems
incompatible with the large size of receptive fields of tectal neurons
Definition: population-coding using mapping combinatorics of
intersecting receptive fields
A non-firing neuron conveys as much information as a firing neuron. All
neurons participate at the information coding
Weakness: likely to suffer from metamery (convergence of information
channels)
 Simulander I



Feedforward network with only 100 neurons, trained by an evolution
strategy for the specific task of head orienting (implies prey localization)
Distribution and sizes of receptive fields of tectal neurons are
respected and firing rates have been adapted
Unstructured mapping: follows the prey with high accuracy
 Simulander II

Similar to Simulander I, but trained for the specific task of frontal
tongue projection (implies depth perception)
Addressed Questions
 How can the stimulus location and depth estimates be extracted from




the tectum maps?
What sensorimotor transformations occur at the level of the optic
tectum, the brainstem and the pathways between them? Can a
structured mapping provide an accurate visual tracking?
Which type of a tectum-brainstem mapping explains the typical curved
approach in monocularized salamanders?
How is the visual perception influenced by head motion during the
approach toward a stimulus? Are additional mechanisms necessary for
dealing with the remaining shifts in the visual background?
Which mechanism implements the release of the snapping behavior?
And how is the tongue controlled?
Neural Networks
 Restrictions (performance motivated):
Uniformly distributed neuron units
 Square receptive fields
 Specification:
 Center receptive field (in degrees of visual field) : determines the
size of the neural network
 Surround receptive field (in degrees of visual field) : determines the
overlap and redundancy feature
 Weights matrix, activation function and thresholds
 Features:
 Reduction (biologically motivated)
 Visualization (extremely practical)

Eyes of the Simulated
Salamander
 Virtual cameras: extract views using provided OpenGL functions
 Correct the view using a spherical projection
 Photoreceptors are equivalent to pixel grey values
 Scalable visual field
Retinal Ganglion Cells of
Type 1
 Properties:
small size excitatory (2-3°) and strong
inhibitory (12-16°) receptive fields
 no response to change in light
 involved in local contrast calculation =>
edge detector
 project to the contralateral spinal cord =>
obstacle avoidance?
 give rise to a fine grained representation of
the visual field in the retina
 Modelled with the laplacian of a gaussian filter:

 x2  y2  
LoG   1 
e
2 
2 

x2  y2
2 2
 Classic edge detector in computer vision and
confirmed by the study of a larval tiger
salamander retina receptive field
Retinal Ganglion Cells of
Type 2
 Type 2 retinal ganglion cells respond only to moving objects => motion
detectors
 Detection of change: compare the corresponding pixels at different
times, using a linear difference function:
 f x, y, j   f x, y, k 
f dif x, y   
0

if f x, y, j   f x, y, k     0
otherwise
where τ is a threshold, j and k are moments in time, x and y are the
pixel positions in the frame
 Biological inspirations:
 Reflects signals with delayed pathways that give rise to a
simultaneous representation of the same object at different times in
the brain
 Flat weights: the activity sharply increases when an object enters
the receptive field variation
 Tectal neurons are contrast-sensitive: linear function
Retinal Ganglion Cells of
Type 3
 Properties:




large receptive fields (10-20°)
tonic response to change in light intensity
respond to overall luminosity (dimming detectors)
respond also at low contrast and velocity
 Model:


flat weights
simple summing network
 Predator detectors among other
Optic Tectum Model
 Principal biological inspirations:
Retinotopic map in the optic tectum: electrical stimulations result
in turning movements that roughly correspond to this map
 Only two synapses between the retina and the brain stem: the
tectum directly projects onto the brain stem
 Input: retinal ganglion cells of type 2. Motion is a necessary
prerequisite for a stimulus to be interpreted as prey
 Structured mapping: different strengths along the rostro-caudal axis,
reflects the stimulus eccentricity
 Motoneuron activation function (integrating weighted tectal activity):

f M     i xij
i
j
where x is the change in light intensity of the pixel at positions i and j
 Linear weights function:
where α and β are parameters
-
     
+
Optic Tectum Model II
 Version with ipsilateral input
(contribution from both eyes)
f M    contra  i xijcontra    ipsi  i xijipsi
i
j
i
j
 ipsi     ipsi   ipsi
 contra    contra   contra
Optic Tectum Model III
 Normalizing tectal activity. The modified model is robust to changes in
the stimulus parameters and visual scene:
fM 
 
i
j
contralateral
i
j
contralateral
ipsilateral
x

x
 ij
 ij
i

i xijcontralateral    ipsilateral i xijipsilateral
j
Biological reference:
 TO4 neurons, arborize in RGC2
 TO2 neurons, large receptive fields
both project to the nucleus isthmi
i
j
Pretectum Model
 Large stimuli: based on RGC3 =>
dimming detectors with three times
larger receptive fields
 Motion: compare direct and indirect
(via tectum) RGC3 responses
 Direction-sensitive neurons:



Why temporonasal sensitivity?
Based upon separating the ON
and OFF channels
Hypothesis: only sensitive to dark
objects (biologically consistent)
Snapping Model
 Relevant for depth estimation
 Tongue mechanism (biologically consistent): 4 muscles, protraction and
retraction times modulated by the stimulus position
 We proposed a mechanism for frontal snapping based upon divergent
projections of the tectal neurons
Results
 Find optimal α and β parameters of the linear weights function through
an exhaustive search of the parameter space
 Cost function: difference between the stimulus direction and the
salamander orienting movement
 Experimental conditions
 Ewert experiment: the stimulus is moved on a semi-circular
trajectory with a constant speed in front of the animal
(task of head orienting)
 Body muscles were inhibited (only neck muscles)
 The stimulus parameters (size, speed, distance, ...) and network
parameters (number of neurons) were fixed according to values
found in literature. Both single stimulus and complex background
were used
Optimal Values
 Many combinations of values give similar results
 Good results are also achieved without ipsilateral input
α and β Parameters
 Regular parameter space. With different fixed values the aspect is
conserved and the minimal error area (in black) is shifted
 β parameters are not essential, the minimal error area is centered in
point (0,0)
α contralateral (x-axis) and α ipsilateral (y-axis)
with optimal β parameters
β contralateral (x-axis) and β ipsilateral (y-axis)
with optimal α parameters
Performance Results
 An accuracy of less than 3° (real value) is achieved for small stimulus
velocity values with 20000 RGC2 and 2000 (less than 3300) tectal
neurons.
 Robustness : stable reaction to change in stimulus parameters and
visual scene
With Complex Background
 The salamander has difficulties with following the prey stimulus as the
amount of “noise” is considerable. It discriminates between objects with
same apparent angular size, however orients at "average flies"
 The model should be coupled with a selective visual attention
mechanism (enhanced retinal signals in the area containing the prey
stimulus) and/or optokinetic or vestibucollic image stabilization reflexes
(antagonistic head movements that compensate for body undulations)
 Integrating approach is trivial with a unique prey stimulus
Pretectum
 The salamander discriminates between a small prey object and a large
predator object
 When the pretectum is abolished, escape behavior fails to occur
 Delayed response: the salamander escapes for a longer time than the
predator is visible
 Weakness: based upon angular size, close prey may be interpreted as
predator. Therefore the threshold is essential (arbitrary as no data
exists on predation)
Snapping
 No additional neurons, based upon divergent patterns of tectal neurons
projections
 Consistent with biological lesion data:
 codes for “closeness”
 realistic precision (about 30%)
 Depends on the movement direction
Reproduced Phenomena
 Lesion and stimulation experiments:
Lesion and stimulation of the optic tectum
 Lesion of the pretectum
 Generation of saccadic movements
 Monocularized salamanders
 Prey preferences

Saccadic Movements
 Pursuit movements such as the head accelerates for a few seconds,
until maximum velocity is reached, and then is released
 We attribute them to the tectal cells resolution
Monocularized Salamander
 With one eye covered, H.Italicus shows a conspicuous approach
behavior toward a prey stimulus. It takes a curved path and bends its
body toward the side of a seeing eye, compensating by turning the
head between 60° and 90°
Monocularized Salamander
Prey preferences
 All preferences are inherent to the network!!
Comparison to
Previous Works
 Simulander I






More neurons, but still biologically plausible (2000 vs. 100)
Less accurate, more realistic (2°-6° vs. 1°)
Inherent preferences vs. a function reflecting the stimulus size and
velocity (corresponds to the observer’s knowledge)
No real distribution of tectal neurons, respected in Simulander
Faster reaction
No positions in which stationery prey elicit orienting behavior
 Simulander II


Lower precision, but more realistic (90%, real success rate 40%)
In Simulander far objects elicit more activity, double inconsistency
(should code for “closeness” and further objects seem smaller)
Response to questions
 Extraction of stimulus localization and depth estimates can be achieved




with a structured mapping between the optic tectum and the brain stem
The sensorimotor transformation of the horizontal angular distance of
the tectum neurons to muscle activity can provide an accurate prey
localization.
Direct observation of the influence of head motion during the approach
is provided. Additional mechanisms for dealing with the self-motion
visual shifts are necessary
The investigated tectum model accounts for the typical curved
approach in monocularized salamanders
A plausible mechanism that acts as a releaser for the snapping
behavior is proposed
Conclusion
 We have implemented models of the three types of retinal ganglion
cells, the optic tectum, the pretectum and a tongue projection
mechanism that account for the typical feeding sequence and escape
behavior
 The optic tectum model reproduces many experimental data
 Everything is observable
 Warning: data and methodology dependant
 Our salamander resembles a newly born salamander thrown in the
world
Future Work










Study a tectum model with nonlinear weights functions
Use the real distribution and receptive fields sizes of tectal neurons
Time-dynamics vs. discrete time steps
Study the effect of overlapping fields (redundancy => error resistant,
maybe emerging properties)
Implement a visual attention model
Implement experience-based models such as habituation
Further development of the pretectum model
Extend the model to other brain areas such as the nucleus isthmi or
thalamus (obstacle avoidance)
Develop a more elaborate model for depth estimation (not only frontal)
Work on an object-background discrimination with respect to selfmotion shifts of the visual input