Navigation 4: Cognitive Maps Example: “Toto” Example: “Toto

Connect by
raw motor
action
Topological Raw sensor Store few
data
locations
Extract
features
Navigation 4: Cognitive Maps
Connect with
some metric
information
Store
continuous
links
Example: “Toto”
• Mataric (1990)
• Robot with compass and 12 sonar sensors
• Subsumption architecture to follow boundaries:
Correct
Make
Nodes have
Use sensor
inferences
global
model
between nodes position
Align
Avoid
Stroll
Metric
Absolute
Convert to Continuous
spatial data representation metric
Example: “Toto”
• Identify ‘wall’, ‘corridor’ or ‘junk’ as dynamic
landmarks (i.e. pattern of sensor input and
movement over time)
LW4
Example: “Toto”
C0
C4
LW0
LW4
LW8
C12
OR
Stop, go, reverse
heading
Motor
output
• Then add layers for landmark recognition, storing,
and navigation
Place-cells
LW6
LW12
• Neurons in rat hippocampus show property of
increased firing rate whenever the rat is in a
particular place in its environment
LW0
C4
LW6
LW8
C0
LW10
C12
• Stored in simple graph structure:
Place-cells
• In a new environment, place cell coding
gradually emerges, then remains consistent
• Same cell may code different parts of space in
different environments
• Neighbouring cells do not necessarily map
neighbouring positions in space
• Encoding will move with visual cues, but still
active in the dark
• Activity over the population of cells can be
used to predict the rat’s position
• Graph reflects the sequence of landmarks
encountered when exploring
• Can find shortest path from present location to
goal by activity propagation
• Draw analogy to rat place-cell system
Place-cells
• There are many models of this system:
– Some are ‘inspired’ by rat but not really relevant to
understanding the biology.
– Vary in the level represented, e.g. algorithms representing
brain area functions vs. neural models.
– Include more or less detail e.g. the number of brain areas
modelled.
– Accuracy is limited by some gaps in current knowledge,
resulting in several inconsistent alternative models.
– Differ in generality i.e. the range of rat (or other animal)
behaviour and neural data explained by the model.
– Also differ in how closely the behaviour is reproduced –
not many provide more than a a qualitative match.
– A number have been implemented using the medium of a
robot to test in real world.
E.g. Burgess et al 1997
•Visual detection of
floor-wall edge coded
by 4x15 distance-tuned
sensory cells.
•900 entorhinal cells
activated by pairs of
sensory cells.
•Place cells selforganise to represent
patterns of entorhinal
response.
•Associates place cell
response with goal
E.g. Arleo & Gerstner 2000
E.g. Arleo & Gerstner 2000
•Visual and path
integration
information
combined to create
place-cell coding
• Unsupervised
Hebbian learning to
organise the spatial
representation
• Each function
mapped to specific
brain areas.
• Pose (position and
orientation) as a competitive
attractor network
• Model EC as two
areas, one receiving
input from head
direction system.
• Packet of activity represents
position and is influenced by
visual and PI inputs.
• Output of
hippocampus to
nucleus accumbuns,
learning is via VTA
dopamine system.
• Explicit area for
navigation with
reward-based
learning
References:
Place-cells
• Rat place-cells are generally interpreted as an internal
spatial representation or cognitive map
• However there is a lack of clear evidence that ‘map
read-out’ to perform navigation occurs:
“Extensive research…has yielded a large database, but not
yet solved the basic question of whether rodents are
capable of true piloting on the basis of a truly cognitive
map” Etienne, 1998
• Most animal navigation can be explained by a
combination of homing, route learning and path
integration.
• Are robots a good model for the rat? Are rats good
models for robotics?
E.g. ‘RatSLAM’ 2004
Maja J Mataric (1990) Navigating With a Rat Brain: A Neurobiologically-Inspired
Model for Robot Spatial Representation in Proceedings, From Animals to Animats:
First International Conference on Simulation of Adaptive Behavior (SAB-90), J-A.
Meyer and S. Wilson, eds., MIT Press, 169-175.
N Burgess, JG Donnett, KJ Jeffery J O'Keefe (1997) Robotic and neuronal simulation
of the hippocampus and rat navigation. Phil. Trans. Roy. Soc., London B 352:15351543
Angelo Arleo and Wulfram Gerstner (2000). Spatial Cognition and Neuro-Mimetic
Navigation: A Model of Hippocampal Place Cell Activity. Biological Cybernetics,
Special Issue on Navigation in Biological and Artificial Systems, 83:287-299.
M. J. Milford, G. Wyeth, D. Prasser (2004) "RatSLAM: A Hippocampal Model for
Simultaneous Localization and Mapping," presented at the International
Conference on Robotics and Automation, New Orleans, United States, 2004
A.S. Etienne (1998) Mammalian navigation, neural models and biorobotics.
Connection Science, 10: 271-289
A.T.D. Bennett (1996) Do animals have cognitive maps? Journal of Experimental
Biology, 199:219-224
R. Wehner(1990) Do insects have cognitive maps? Annual Review of Neuroscience,
13:403-413.
• Tested in real environments
with some success
• Point at limitation that rat
system cannot maintain
multiple pose hypotheses