Discovery and Neural Computation

1
Discovery and Neural Computation

• Paul Thagard
• University of Waterloo

2
Outline

• Discovery
• Neural Computation
• Multimodal representation
• Abduction
• Conclusions

3
Creative Scientific Discoveries

• New hypotheses, e.g. sound is a wave
• New concepts, e. g. sound wave
• New instruments
• New methods

4
Why Neural Computation?

• Scientists have brains
• Brains have powerful computational capacities
• Multimodal representations sensory, motor,
  emotion
• Understanding of causality
• Parallel constraint satisfaction
• Cognitive science studies mechanisms at multiple
  levels social, psychological, neural, and
  molecular.

5
Theoretical Neuroscience

• Beyond connectionism, PDP
• Spiking neurons
• Large populations
• Multiple, organized brain regions
• Representations tied to sensory, motor, emotional
  regions

6
Representation

• Neural populations represent the world by
  encoding inputs from external sensors.
  Eliasmith causal correlations.
• Neural populations represent the body by
  encoding inputs from internal sensors.
• Neural populations represent neural populations
  by encoding inputs from neural populations.

7
Neural Representation
8
Multimodal Representations

• Sensory concepts are patterns of firing
  activity in multiple brain regions, e.g. visual,
  auditory, tactile
• Causality sensory-motor-sensory patterns
• Emotions patterns include ones for bodily input
  and cognitive appraisal in regions such as the
  nucleus accumbens
• Emotional consciousness Google Thagard

9
Emotions in Scientific Thinking
beauty happiness
happiness hope
happiness surprise
interest curiosity wonder
Generate questions
Try to answer questions
Generate answers
Evaluate answers
fear anger frustration
avoid boredom
worry
disappointment
10
Hypothesis Generation

• Causal
• Creative
• Simplest form, abductive
• Why effect?
• If cause then effect.
• So maybe cause.

11
Neurocomputing Problems

• How to represent causal if-then?
• How to connect with emotions?
• How to make inference of cause from effect?
• First attempt Thagard Litt, in press.

12
Representation

• Represent if-then relations by holographic
  reduced representations (Plate)
• Relations are vectors built out of vectors for
  concepts and roles
• Translate vectors into neural populations
  (Eliasmith). 6000 neurons
• Simplify emotions as vectors (Litt)

13
Processes

• Representation of B marked as emotionally
  surprising.
• Retrieve A -gt B from memory of rules.
• Extract A by decomposing holographic
  representation of A -gt B.
• Mark A as emotionally satisfying.

14
Current Work

• Understand causality as intervention
• Not just statistical or universal
• Causes make things happen
• Babies and monkeys understand causality
• Sensory-motor-sensory schemas
• Now developing model using Neural Engineering
  Framework

15
Open Problems

• Generation of new concepts
• Not just learning from examples
• Need new nodes based on new experiences
• Theoretical concepts combine previous concepts,
  but how does this work neurally?
• More complex hypothesis formation
• Integration with analogy

16
Conclusions

• Creative discoveries are made by human brains.
• Brains have representational and computational
  resources not present in current AI models, e.g.
  emotion.
• Neurocomputational models of discovery can be
  developed.