A quick introduction to connectionist cognitive architecture - PowerPoint PPT Presentation

1 / 37
About This Presentation
Title:

A quick introduction to connectionist cognitive architecture

Description:

Better than the old 'info processing' model - allows easy ... a photograph, jaggies, ... Example: looking at a photograph (low presence) A. Flatness, ... – PowerPoint PPT presentation

Number of Views:108
Avg rating:3.0/5.0
Slides: 38
Provided by: daven2
Category:

less

Transcript and Presenter's Notes

Title: A quick introduction to connectionist cognitive architecture


1
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

2
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis

3
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis
  • nodes are organized hierarchically in layers,
    increasing in abstractness

4
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis
  • nodes are organized hierarchically in layers,
    increasing in abstractness
  • nodes have a state and trigger point, above which
    they influence other nodes

5
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis
  • nodes are organized hierarchically in layers,
    increasing in abstractness
  • nodes have a state and trigger point, above which
    they influence other nodes
  • nodes tend to excite those in other layers, and
    inhibit those in the same layer

6
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis
  • nodes are organized hierarchically in layers,
    increasing in abstractness
  • nodes have a state and trigger point, above which
    they influence other nodes
  • nodes tend to excite those in other layers, and
    inhibit those in the same layer
  • on a stimulus, activation spreads through the
    network until it settles into a new state

7
A quick introduction to connectionist cognitive
architecture
  • Simply a toolbox for creating psychological
    theories
  • Better than the old info processing model -
    allows easy feedback parallelism

Assumptions Used
  • The mind a series of highly interconnected
    nodes, each encoding a hypothesis
  • nodes are organized hierarchically in layers,
    increasing in abstractness
  • nodes have a state and trigger point, above which
    they influence other nodes
  • nodes tend to excite those in other layers, and
    inhibit those in the same layer
  • on a stimulus, activation spreads through the
    network until it settles into a new state
  • there is a fixed amount of activation available
    to excite nodes (attention)

8
An example of a node
It has wings State 0.4
9
An example of a network
10
Highly useful psychological modeling tools
  • Explain attention findings (divided vs. focussed
    attention)
  • Explain differences between STM and LTM
  • Explain forgetting
  • Explain learning application of concepts
  • Explain priming effects in cognition generally
  • Explain recognition of recall failure

11
Return to presence - the problem at hand
Current approach to understanding
presence identifying variables via empirical
manipulation
For this to work, need to test as many variables
in the same experiment as possible
This blind approach gives awesome practical
problems!
12
Imagine want to test the effects of these
variables
  • Stereopsis (yes/no)
  • Avatar animation (yes/no)
  • Field of view (16/ 32/ 64)
  • Display type (MHD/Fishtank/Desktop)
  • Perspective (1st person/3rd person)

This requires 2x2x3x3x2 72 conditions At a
minimum of 12 subjects per condition, 864
subjects!!
13
Apart from this, current research suffers from
  • Lack of statistical power (small effects, few
    subjects)
  • Little methodological sophistication (still use 2
    group experiments)
  • Conclusions which go beyond data found
  • Lack of understanding of psychological processes
    of perception

14
Solution to the problem Follow standard natural
science methods
  • Begin with a model
  • Generate specific hypothesis implied by the model
  • Test these hypotheses to evaluate the validity of
    the model

Advantage of this approach - there is always a
defined research goal and avoids many null results
15
Current model of presence
Nothing explicitly or formally defined
Model seems to be Being there is a
consequence of having sensory stimulation
approaching that found in natural perception.
Problems with this model
  • No account of higher-level processes (willing
    suspension of disbelief)
  • Gives a technology-centered view of a
    psychological process
  • It is difficult to understand the implications of
    such a model
  • Current evidence contradicts it

16
Towards a new model considerations
Things we know about presence, supported by
current model
  • Negative correlations
  • Bad interface
  • Attention on the RE (BIPs)
  • Poor immersion
  • Positive correlations
  • Realism
  • Attention on the VE
  • Storyline

Most trustworthy results are relational (biggest
effects) No information on causality with
relational studies
17
  • From those findings, several things become clear
  • Direction of attention is important
  • Quality of stimulus is important
  • Interference with the process is possible
  • Top-down processes are important
  • The current presence model does not take many of
    these points into account
  • cannot explain why BIPs occur, or why
    interference should happen
  • does not give any space for top-down processing
    (such as storyline)

18
  • Connectionist networks are a possible candidate
    for modeling
  • they model attentional processes well
  • they can account for interference
  • they can explain selection between competing
    stimuli
  • they can account for top-down effects, bottom-up
    effects as well as their interactions

19
Presence as a form of environmental perception
reaction
In my model, presence is a natural process, a
part of perception can occur in any setting,
not just computer created VEs (theme parks,
books, films, etc)
Rather than thinking of presence as a feeling of
being there, look at it from a behaviorist
view During the presence state, subjects are
more likely to think and act in a way coherent
with the demands of the virtual world rather
those of the real world. The dominance of
virtually-aligned cognition represents the
level of presence.
This idea agrees with the classical being there
idea of presence, but is more operationalized -
bridges presence and behavioral presence
20
My idea of a connectionist model for
environmental perception
Conceptual layers
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Perceptual layers
21
The center layer (layers) represent them
mechanisms which determine presence
These nodes become activated when the subject is
stimulated by something which is perceived as
real (eg. animation, stereo, etc)
Detectors for real stimuli
These nodes are activated when the subject is
stimulated by something which is perceived as an
image (eg. glint of a photograph, jaggies, etc)
Detectors for image stimuli
These nodes are activated when the subject is
stimulated by other stimuli (not emanating from
the VE)
Detectors for other stimuli
As usual, there is vertical excitation and
lateral inhibition
22
The high presence situation Stable state with
Real cluster most activated
Conceptual layers
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Perceptual layers
23
Example looking at a photograph (low presence)
Conceptual layers
A
I am holding a photo, etc.
I
I
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
A
Flatness, glossiness, etc.
Perceptual layers
Result action/thoughts are in terms of an image
rather than an object
24
A photo is a simple case - the Image perceptions
outweigh the Real perceptions
In a VE, the Real, Image and Other clusters are
competing far more strongly
However, in order for a stimulus to affect our
actions and thoughts (presence), only one of
those middle clusters can remain activated enough
to affect the upper layers
Consider a good VE high immersion (HMD,
earphones), real walking to move
25
Example High immersion VR (high presence)
Initial conditions
Conceptual layers
I am in a VE
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Jaggies, low FOV, etc.
Perceptual layers
26
Example High immersion VR (high presence)
Stable state
Conceptual layers
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Perceptual layers
Result action/thoughts are in terms of the VE
rather than the RE
27
This is still a more simple case - what about
presence with very poor immersion equipment?
(Counter Strike presence)
In these cases, top-down processing becomes far
more important - the storyline finding comes into
effect
Consider a bad VE low immersion (desktop), but
player keen to be a terrorist (willing
suspension of disbelief)
28
Example Low immersion VR (high presence)
Initial conditions
Conceptual layers
I am in playing a game
I am a terrorist
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Jaggies, flatness, etc.
disctractions
Perceptual layers
29
Example Low immersion VR (high presence) Stable
state
Conceptual layers
Detectors for real stimuli
Detectors for image stimuli
Detectors for other stimuli
Perceptual layers
Result action/thoughts are in terms of the VE
rather than the RE (moderate amount)
30
Evaluation of the new model - explaining
acquisition of the presence state
  • Strengths
  • Can explain high presence levels in low immersion
    situations
  • Can explain low presence in high immersion
    situations
  • Agrees with the holodeck idea (lucid dreaming)
  • Supersedes and expands the previous model
  • Weaknesses
  • Vague about the processes in the conceptual
    layers (key to presence)
  • Division of middle layer into three clusters
    seems abitrary
  • The other cluster is a bit of a kludge to
    explain distractions

31
Further explanations allowed by this model -
failures of presence
Failure due to interference - distractions High
levels of activation of the other cluster will
inhibit activation in the Real cluster
These distraction can come from below
(extraneous sounds etc.) or from above (dual
tasks eg. BIP detections) Distractions can
block presence, or merely reduce it This
explains the lower presence power of low
immersion systems
32
Failure due to poor stimulus quality The VE
stimuli are always occurs in the presence of
other stimuli For the Real node to become
dominant , the Real stimuli must be high
enough to overcome Image and Other if the
stimuli is too poor, Image will inhibit Real
too much Poor stimuli do not ensure zero
presence, but certainly lower levels
33
Failure due to non-compliant subject This is a
more complex situation - involves 2 simultaneous
failures The first is interference - a
non-compliant subject will be attending to
personal thoughts, activating the Other cluster
from above The second is a lack of top-down
activation of the Real node. This will make it
harder to make it dominant
34
Predictions from the connectionist model
  • Many predictions can be directly made from this
    model, such as
  • A low immersion, poor quality system can still
    create high levels of presence - prime the
    subject, reduce distractions
  • Task performance in VR will be increased in high
    presence situations only if the task requires
    thought in terms of the VE (spacial or thematic)
  • Conversely, these translate into recommendations
    which we can make to VR authors
  • Give more importance to the subjects mental set
    (priming materials)
  • Ensure that immersion is increased, even if
    stimulus quality is low
  • Tap into users previous experiences where
    possible (increase top-down activation)

35
Empirically testing the model
Key statement of the theory Presence levels
are determined by stimulus quality, priming,
distractions all singly as well as in
interaction This should be directly tested, as
it is the foundation of the theory
One possibility - use a 3-way factorial design
(allows checking main effects and interactions)
36
Set up a 3x2x3 design (about 50 subjects if we
use repeated measures - not ideal) Variable 1
stimulus quality - 3 levels Stereo, textured,
radiosity Mono, textured Mono, flat
shading Variable 2 priming - 3
levels preparatory video/booklet/briefing no
preparation Variable 3 distraction - 3
levels no distraction, visual or
auditory infrequent, slight distractions frequen
t intense distractions
37
  • Should find
  • Each variable makes a difference to presence
    levels individually
  • Interactions between variables (eg. high presence
    when priming was high but only if distractions
    were low as well)
Write a Comment
User Comments (0)
About PowerShow.com