Integrating and Controlling Cognition with a Focus of Attention

1
Integrating and Controlling Cognition with a
Focus of Attention

• Nick Cassimatis

2
Reasoning, control and integration

• Domain-general or statistical strategies cannot
  explain all quick adaptation.
• E.g.
• McDonalds arches in Saratoga are pink ? look
  for big pink objects.
• Probably not innate need no trial and error.
• 100 200 shows hints of planning.
• Hypothesis
• Reasoning and planning strategies important for
  control and integration.

3
Integration problem

• Problem How does higher-level cognition
  integrate with lower-level cognition.
• Dont these involve memory, mental imagery,
  attention, too?
• What are mental models?
• Bayes Nets vs. everything else?
• Not clear how search, BNs, rules, neural
  networks, etc. can be integrated into one system.

4
Main ideas

• There is a cognitive focus of attention
• Not tied to a single sensory modality.
• Integrates modules encapsulating lower-level
  cognitive processes.
• Mechanisms for controlling visual attention ARE
  higher-level reasoning.
• This enables more integrated cognitive models.
• Polyscheme cognitive architecture embodies these
  ideas.

5
Desired input

• Up to now, just happy to have integration.
• Now would like to more closely integrate these
  ideas with what is known empirically.
• Empirical evidence for non-perceptual focus of
  attention.
• Evidence either way of visual and cognitive
  attention being the same or different mechanisms.
• Additional attention control mechanisms that are
  relevant.

6
Mind has specialized modules

• Specialized modules, called specialists, using
  different mechanisms.
• Spatial memory grids.
• Physical motion prediction rules.
• Object recognition neural networks.
• Etc.
• Agnostic on perceptual or motor basis.
• Generalized mental imagery specialists can
  represent past, future, hypothetical, etc. worlds.

7
Cognitive focus of attention

• Generalize evidence for integrative visual focus
  of attention (e.g., Stroop, dual-task and visual
  search)
• Cognitive focus of attention.
• All specialists focus on the same things at the
  same time and make inferences about it.
• Evidence these extend to higher-level cognition
• Semantic interference Semantic/emotional, Stoop
  effects.
• ?

8
How is cognitive attention controlled

• Suggestions from visual attention
• Habituation
• Dont focus on something for too long.
• Negative priming
• Suppress distractors.
• Probability
• Focus on most likely outcomes.
• Change
• Focus on changing aspects of scene.

9
Main point

• These mechanisms applied to cognitive focus of
  attention with imagery implement much human
  reasoning.

10
Backtracking Search

• Choose A or B?
• Focus on world where B is taken because of
  probability.
• Inhibit A because of negative priming.
• Stop focus on B because of habituation.
• Focus on world where A is taken.
• Focus on world where A and then C is taken
  because of probability.
• Focus on world where A, C and then GOAL is taken
  because of probability.

11
Backtracking Search

• Order of foci same as that in backtracking
  search.
• Strength and duration of negative priming
  determines whether depth-first or breadth-first
  search.

12
Other algorithms

• Stochastic simulation (for Bayesian inference)
  implemented by focusing on more probable outcome.
• When A is more likely than not-A, focus on the
  world where A is true more often to not-A in
  proportion to how much more likely it is.
• P(A) of world where A is true / of worlds
  where A is false
• Truth-maintenance by focusing on change
• When your belief about A changes, focus on all
  the evens that involved A.

13
Recognizing problem 100200

• Lookahead (in imagery) with brute force strategy
  because of frequency.
• Habituation tires of that.
• Only addition
• Imagery.

14
Attention control is reasoning

• Attention control mechanisms guiding a cognitive
  focus of attention implement reasoning
  algorithms.
• Execution of a reasoning strategy is a set of
  attentional fixations
• F1, F2, , FN
• No need to posit a reasoning module, at least for
  a lot of reasoning.

15
Explanation of integration

• High-level and low-level cognition
• Reasoning strategies executed as sequence of
  fixations.
• Each fixation involves all of the lower-level
  specialists.
• High-level reasoning strategies with each other
• Reasoning strategies executed as sequence of
  fixations.
• Interleaving these is very easy.

16
Integration of perception, memory, attention and
reasoning
17
Explaining integration of high-level cognitive
processes with each other.
18
Polyscheme
19
Compare and contrast

• Borrows from existing modeling frameworks.
• Worlds Search, (Bayesian) stochastic simulation,
  Soar, Johns-Laird mental models, situation
  semantics.
• Impasses Attention control strategies initiated
  during conflict.
• Attention control ACT-R.
• Differences
• Designed from the ground up allow many
  representations as first-class citizens.
• Higher-level of basic services.
• Focus on language.

20
Why Polyscheme

• Enables models that combine aspects of existing
  cognitive architectures.
• E.g., physical reasoning model
• Neural networks for object recognition.
• Rules for physical dynamics.
• Impasses, simulation of states and
  activation-based conflict resolution for
  reasoning about uncertainty.
• Search.
• Stochastic simulation.
• Truth-maintenance for belief revision.

21
Existing Polyscheme Models

• Used at NRL, RPI and (soon) AFRL.
• Models
• Infant physical reasoning.
• Syntactic understanding.
• Uses same mechanisms as physical reasoning model.
• Human-robot interaction.
• Goal
• Only one Polyscheme model.

22
Conclusions

• Attention control is reasoning.
• Enables integration of reasoning and problem
  solving modeling with memory, attention,
  perception and action modeling.
• Reasoning and problem solving strategy execution
  as sequence of fixations key to integration.
• High-level integrated with each other through
  interleaved fixations.
• High-and low-level integrated since each fixation
  involves all the low-level specialists.
• Attention is not a subsystem or module, but
  fundamental organizational mechanism of cognitive
  architecture.
• Embodied cognition not in conflict with
  old-fashioned reasoning and problem solving
• If you have a system with imagery and focus of
  attention, then if you control that focus of
  attention, you are reasoning.