Roundtable Discussion

1
Roundtable Discussion

• Bradley Love

http//love.psy.utexas.edu/
2
Differing theories ordiffering frameworks?
3
What constraints are offered?

• Barring minor linguistic relativity, you can
  express the same thing in Italian or English.
• Of course, some languages are better at
  expressing certain things.

4
(No Transcript)
5
(No Transcript)
6
The Connectionist Model
7
The Connectionist Model

• Response competition between previous experienced
  patterns (bias) and the current pattern.
• A parameter governs the relative strength of
  these forces.

8
Dynamical Systems Model
9
Dynamical Systems Model

• Response competition between previous experienced
  patterns (bias) and the current pattern.
• A parameter governs the relative strength of
  these forces.

10
Frameworks as Guiding Metaphors
11
Frameworks as Guiding Metaphors

• Connectionism focus on learning, representation,
  and mechanism.
• insights into brain function, PFC development.

12
Frameworks Offer Guidance

• Connectionism focus on learning, representation,
  and mechanism.
• insights into brain function, PFC development.
• scaffolding and generalization explanations.
• Dynamical systems focus on coupling of memory,
  body, environment, perception, and action through
  time.
• clever experiments on posture, wrist weights,
  gaze/reach.

13
Other frameworks?

• Good old fashion cognitive psychology

14
Other frameworks?

• Good old fashion cognitive psychology
• probability learning
• recency effects
• negative recency effects
• sequence learning
• artificial grammar learning
• SRT task

15
Probability Learning
16
People Resonate to Patterns
17
People Resonate to Patterns
18
Linear Associative Shift-Register (LASR)
19
Sequential Predictions
20
Sequential Predictions
21
Linear Associative Shift-Register (LASR)
22
Combining Connectionist and Dynamical Systems
Approaches

• Reinforcement Learning
• learning and performing actions through time.
• maximizing discounted reward through time.
• driving a car, navigating, walking, playing
  backgammon, flying a helicopter, giving a lecture.

23
Reinforcement Learning

• Involves learning Q-values and following a policy
• state action pairs What action should I perform
  in this state?
• Connectionist models learn the state action
  pairs.
• low dimensional representation that is
  continuously updated.
• Dynamical systems are reflected in the policy.
• what should I do NOW following the markov
  condition.
• also informs what should be included in the
  state.

24
Reinforcement Learning

• Involves learning Q-values and following a policy
• state action pairs What action should I perform
  in this state?
• Connectionist models learn the state action
  pairs.
• low dimensional representation that is
  continuously updated.
• Dynamical systems are reflected in the policy.
• what should I do NOW following the markov
  condition.
• also informs what should be included in the
  state.
• Operating at different time scales.

25
OK
Lets Talk.
26
Some Questions

• Are we arguing about frameworks or theories?
• Would other frameworks be helpful?
• Can we combine these frameworks?

27
Some Questions

• Are we arguing about frameworks or theories?
• Would other frameworks be helpful?
• Can we combine these frameworks?
• Stability vs. Instability