Computational Cognitive Models of Spatial Memory in Navigation Space: A review

Computational Cognitive Models of
Spatial Memory in Navigation Space:
A review
Madl et al
2015
Neural Networks

Index
 Introduction
▫ Spatial memory and representations
 Computational cognitive models of spatial memory
▫ Symbolic spatial memory models (in real-world and simulations)
▫ Neural network-based spatial memory models (in real-world and
simulations
▫ Spatial memory models in cognitive architectures
 Conclusion
 Discussion

Speed cell [3]
Neural Spatial Representation
 Egocentric spatial memory
▫ Egocentric representations represent spatial information relative to the
agent’s body or body parts
Angular head velocity cell [1] Head pitch cell [1]
[2]
[1] Stackman and Taube, 1998, J of Neurosci. [3] Kropff et al., 2015, Nature
[2] Bassett and Taube, 2001, J of Neurosci.

 Allocentric spatial memory
▫ Allocentric representations represent spatial information relative to
environmental landmarks or boundaries, independent of their relation to the
agent
[4]
[4] Yoder et al., 2011, Trends in Neuroscience

 Transformation to spatial memory
▫ Egocentric and allocentric representations are driven by sensory information
[5] Kravitz et al., 2011, Nat Rev Neurosci.
[6] Hitier et al., Front. Integer. Neurosci., 2014
LMN
MEC
HP
AD LD
PaS PoS
DTN
MVN
[6]
HP : Hippocampus
MEC : Medial entorhinal cortex
PaS : Parasubiculum
PoS : Postsubiculum
AD : Anterodorsal thalamus
LD : Laterodorsal thalamus
LMN : Lateral mammillary nuclei
DTN : Dorsal tegmental nucleus
MVN : Medial vestibular nuclei
Place cell
Border cell
Grid cell
Head direction cell
Angular head velocity cell
[5]
Posterior parietal cortex
Retrosplenial cortex

Spatial Learning and Memory
 Hebbian learning theory in spatial memory
▫ Fire together, wire together
▫ Unsupervised learning rule
▫ Hebb’s rule is critical for associative learning between the representation of a rat’s
current location, and sensory stimuli at that location
Hebbian learning
Head direction
Distance from boundary
Smell
Visual cue
Light
Sensory stimuli

Spatial Learning and Memory
 Reward-based learning theory in spatial memory
▫ Supervised learning rule
▫ The dopamine system
Place cells (n = 12)
[7]
[7] Hok et al., 2007, J of Neurosci.

Computational Cognitive Models to spatial memory research
 Symbolic spatial memory models
▫ Models emphasizing explicit rule and localist representations based on symbolic logic
▫ Discrete states
 Neural network-based spatial memory models
▫ A number of simple processing units affecting each other via weighted connections
▫ Operate in parallel
▫ Commonly learn rules from training data instead of explicit rules
▫ Not discrete
 Cognitive architectures

Symbolic Spatial Memory Models
 The key ideas of symbolic spatial memory models are:
▫ “To make a map and place tree” – save the feature information of place
▫ Construction of represented specific place
▫ Absolute space representation like a map (ASF)
▫ Tree of consistent topology of all places
 Main methods are:
▫ Memory for the immediate surroundings (MFIS)
▫ Distance sensor
▫ Camera

Symbolic Spatial Memory Model
 Models evaluated in real-world environment
▫ 10 laser rangefinder sensor around robot body
▫ To get the boundary element as a represented place
[8] Jefferies and Yeap, 2008, Springer Verlag
[8]

▫ Based on Simultaneous localization and mapping (SLAM)
▫ Laser rangefinder sensors
▫ Place tree
[9] Beeson et al., 2010, The international J of Robotics Research
[9]

[10] Franz et al., 2008, Springer
[10]

 Models evaluated in simulation
▫ Based on semantic network
▫ Artificial intelligence (AI) approach
[11] Brom et al., 2012, Cognitive Systems Research
[11]

Neural network-based spatial memory models
 The key ideas of network-based spatial memory models are:
▫ Non-local (not discrete) and distributed representations
▫ Neural network (NN) are simplified models of the brain composed of a
number of units with weighed connections between them
▫ To make biological plausible (biological NN, BNN)
▫ Artificial neural network (ANN) as a perceptron (input layer, hidden layer,
output layer)
▫ Spiking neural network (SNN) that consists of spiking neural model (e.g.
Integrated-fire model, Izhikevich model, Hodgkin-Huxley model)

Neural network based spatial memory model
 Model evaluated in real-world environment
Sensory cells represent
the coordinate of environment
with allocentric direction
▫ Hierarchy model
▫ Goal-directed task
• ANN model (Gaussian distribution)
• Using camera to detect the
boundary of environment (sensory
cells)
• The place field is constructed by
combination of sensory cells fire
• When the agent locate in goal (pre-
defined), a goal cell is exited
• The activation of goal cell will
encode the proximity to the reward
(Controller cell)
Hard wired
[12]
[12] Burgess et al., 2000, Biological Cybernetics

▫ Hierarchy model
▫ Navigation and map learning
• ANN model (rate coded)
• Using panoramic camera to
detect the place and turning angle
using Gabor filter
• The obstacle avoidance by
proximity sensors
Local view
Allocentric
place cell
Position
integrator
Combined
place code
Nucleus accumbens
Dead
reckoning
Step cell
similar column
Rotation cell
Ation cell
Hebbian learning
[13]
[13] Strosslin et al., 2005

▫ Hierarchy model
▫ Shortest path learning
• ANN model
• Using camera to detect the
landmark (orientation and distance)
• The specific place will be
transposed as a node
Possible action
Trial 1 Trial 4
Locomotion
Camera input
[14]
[14] Barrera et al., 2011

 Model evaluated in simulation
▫ To make a hierarchical cognitive
map
▫ Navigation and map learning
• ANN model
• Three different networks
representing associations between
all landmarks
• Step 1: exploration stage for
building level 1 network
• Step 2: building a hierarchical
network based on the distance
from each landmarks
Each nodes represent a landmark
Planning path
Landmark
[15]
[15] Voicu, 2003

▫ To suggest the mechanism of
path integration system
• Cyclic attractor map (CANN)
model
• Each position layers have vector
field according to exited direction
Head
direction
Angular
velocity
Connection
Position
(place cell)
2D CANN
Feedback
Direction
Sensory
input
Locomotion
[16]
[16] McNaughton et al., 1996

▫ To investigate the transformation
between egocentric information
and allocentric information
Coordinate of agent
Direction and distance
Coordinate of agent
perirhinal
parietal window
Various
synaptic weight
[17]
[17] Byme et al., 2007

[18]
[18] Erdem and Hasselmo, 2012

Cognitive architecture based spatial memory models
 The key ideas of cognitive architecture based spatial memory
models are:
▫ Cognitive models of specific processes can be implemented within the
framework of a system-level cognitive architecture
▫ Providing detailed, formal explanations, providing hypotheses and guiding
research
▫ Symbolic, NN or hybrid
▫ Adaptive control of thought rational (ACT-R): “IF”  “THEN” based on
declarative memory (encoding factual knowledge, chunks)
▫ Updating the ACT-R network based on learning

1. ACT-R request is occurred
(e.g. finding cityhall)
2. The request is compared with memory
3. The request is compared
with subnet
(e.g. in the park)
4. “THEN” is created
(e.g. 1150 bus, subway)
5. Select the best “THEN”
[19]
[19] Schulthesis and Barkowsky, 2011

▫ Hierarchical cognitive map
▫ Learning intelligent distribution agent (LIDA) consist of three phase
• Understanding
◦ Sensing environment
◦ Detecting features
◦ Recognizing object and categories
• Attending
◦ Attend to portion of representational environment
• Action
Sub map
Reference frame Sub map
Object
[20]
[20] Madl et al., 2013

Discussion
 The cognition model with biological plausibility is important to
understand how brain works
 However, models with highly complex real-world environments
usually have trouble handling highly complex real-world
environment
 Models built to work well on real-world robots usually cannot be
called biologically realistic, and also have difficulties fitting
behavior data of human or animal
 Therefore, it is very difficult to implement and run a model that
incorporates both high psychological and biological plausibility
and the ability to handle real-world environment

Discussion
 The symbolic based models are constructed with functional
modules and intuitive flow algorithm
 The NN based models are constructed with distributed system
Method Strength Weakness
Symbolic based
Intuitive flow algorithm
Simple algorithm
Fast tree search
Modularization
Biological plausible
High performance
Huge memory space
Low flexibility
Neural Network
Biological plausible
Distributed system
No memory space
High flexibility
High performance
Hard to get a source of
result

Discussion
 The previous symbolic and NN model for spatial memory is top-
down approach
 Even the NN model was constructed based on biological plausible
hierarchical network, they could not represent the bottom-up
cellular mechanism for spatial memory
PoS
PoS
MEC
CA1
SUB
NA

Computational Cognitive Models of Spatial Memory in Navigation Space: A review

More Related Content

What's hot (14)

Similar to Computational Cognitive Models of Spatial Memory in Navigation Space: A review (20)

More from Seonghyun Kim (13)

Recently uploaded (20)

Computational Cognitive Models of Spatial Memory in Navigation Space: A review