Every second counts: Real-time dynamics constrain what is

Document Sample
Every second counts: Real-time dynamics constrain what is Powered By Docstoc
					      Every second counts:
 Real-time dynamics constrain
what is learned and what develops

          John P. Spencer
         University of Iowa
Prof. Dr. Gregor Schöner
& Claudia Wilimzig
Institut für Neuroinformatik
Ruhr-Universität                               Larissa Samuelson
Bochum, Germany                                  University of Iowa
                                                    Anne Schutte
Linda Smith          Alycia Hund                      University of
Indiana University Illinois State U.
                                                  Nebraska-Lincoln
The SPAM Lab…
   Vanessa
   Simmering                       John
                                   Lipinski
   Jeff
   Johnson                         Brandi
                                                                Wendy
                                   Dobbertin
                                                                Troob
   Numerous undergraduate RAs!
 Developmental Psychopathology

Dynamic systems theorists Connectionists tend to
 tend to have OCD           have ADHD
 Obsession with             A somewhat impulsive
 measurable catastrophe     proliferation of ideas and
 flags                      models
 Obsession with formal      Attentional capture such that
 dynamical concepts         every phenomenon is
 Obsession with second-     interesting (we can model
 to-second details of       that!)
 behavior, integration of   Impressive level of
 time scales, embodiment    production (or hyperactivity)
  Developmental Psychopathology
Different [psychopathologies] is good (Smith &
  Samuelson

An exciting future lies in integration, but we must learn
 to cope with one another’s pathologies…
 DFT folks need to learn how to lay some ideas out
 there, even when they haven’t fully ―cooked‖ and
 appreciate the insights that come from public
 exploration of ideas
 Connectionists have to be willing to slow down a bit,
 tolerate some obsessive use of concepts, and work
 toward a thoughtful integration of ideas that we both
 agree will advance our science
 Developmental Psychopathology
I agree with Denis—integration is absolutely the
   correct idea and well within reach…
   Any perceived resistance to that is just my
   OCD talking…

   My name is John Spencer, I’m a dynamical
       systems theorist, and I have OCD.
  Developmental Psychopathology
Goal of today’s talk is to move toward ―normality‖ by
 laying some cards on the table
 What can we capture with the DF approach including
 (dare I reveal) what is on the horizon?
 We do capture learning and we do capture
 development, but in a way that always reveals new
 questions (which gives people wonderful chances to
 critique us…[one origin of our OCD?])
 This makes people feel unsatisfied because we are
 trained (despite our attempts at un-training) to think
 in terms of primary causes and ―mechanisms‖
 Perhaps an understanding of development requires a
 new mind-set
            Dynamic Balance
From a dynamic systems perspective, behavior
reflects a delicate balance among stability, instability,
and flexibility
– Stability is critical for generating goal-directed behaviors
  (e.g., distractor suppression, on-line updating)
– Instability is critical for sensitivity to context and for
  probing the range of possible behaviors (e.g., ―distractors‖
  can indicate that a current goal should be de-stabilized)
– Flexibility reflects a dynamic balance between stability and
  instability (flexibly move from one stable behavior, through
  an instability, to another stable behavior)
Learning and development are about changing the
system’s dynamics to adjust this balance
               Overview of Talk
The focus today is on how this dynamic balance is
  realized within the Dynamic Field Theory
  Part 1: describe this dynamic balance within the
  context of the rich, real-time dynamics of the DFT
  Part 2: focus on longer time scales, illustrating our
  approach to the integration of time scales
   – Real-time dynamics constrain what is learned
   – Real-time & learning time dynamics constrain what
     develops
  General Theme: to highlight the strong interplay
  between theory and experiment in the DF approach
   – Offers constraints and evidence for theoretical concepts
   – It also produces OCD and many sleepless nights!
            Real-time Dynamics
Empirical examples are from relatively simple cognitive
 tasks with three special characteristics…
 Working memory tasks: this allows signatures of
 real-time dynamics to shine through because behavior
 is not dominated by input-output relations
 Metrics: probing memory for items along continuous,
 metric dimensions (not just yes/no tasks) to reveal
 interactions among the many contributors to behavior
 Time: probing systematic changes in memory over
 carefully controlled short-term delays (e.g., 0 – 20 s)
                  Metric Working Memory Tasks
                        2000 ms                             10 sec delay                  Ready, Set, Go!

                 -40°




                             +                                          +                        +




                 Color Space               Target appears            Mouse Response
                     0°
                                                                       Rotated 110° arc


                                          -70°
-90°                             90°
                                         Target                  -70° Color
                 -70°
                                                                  at -35° in   +
                                                                 task space
                  +/-180°
 Error (º)




                        Close Colors (N = 12)               Far Colors (N = 12)
             5                                    5
                                 -110º
Dynamic Field Theory: Overview

                      Perceptual Field                            10
                                                                       center
                                                                  8    20 degree




                                       mean constant error (°)
       ∆y                                                              40 degree
                                                                  6
                          Excitatory                                   60 degree
                                                                  4    80 degree

                      Spatial Working                             2


                       Memory Field                               0

                                                                  -2

                         Inhibitory                               7    0

                                                                  6




                                        mean variable error (°)
                         Excitatory                               5

                                                                  4

                        Long-Term                                 3

                                                                  2
                       Memory Field                               1


                         Inhibitory                               0
                                                                       0
Working Memory For Locations

                                                            10
                       Target Input                         8
                                                                 center
                                                                 20 degrees




                                 mean constant error (°)
     ∆y                                                          40 degrees
                                                            6
                                                                 60 degrees
                                                            4    80 degrees

                      Self-sustaining                       2


                        WM peak                             0

                                                            -2
                                                            7    0        5

                                                            6
                      Peak is stably




                                  mean variable error (°)
                                                            5


                     self-sustaining                        4

                                                            3

                      but contents                          2



                       “drift” over
                                                            1

                                                            0
                                                                 0        5

                          delay
                              Evidence for delay-dependent spatial drift
                                         Participants                        Dynamic Field Model
                          10                                           10
                                center
                                20 degrees
                                                                            Time-dependent spatial drift
                          8                                            8
mean constant error (°)




                                                                               away from midline
                          6                                            6

                          4                                            4

                          2                                            2

                          0                                            0

                          -2                                           -2
                                0        5         10       15   20           0     5       10       15   20
                          7                                            7

                          6                                            6
                                                                                        Stability at midline
mean variable error (°)




                          5                                            5

                          4                                            4

                          3                                            3

                          2                  Variance grows in time    2

                          1
                                             (change in stability is   1

                          0
                                             the correct picture)      0
                                0       5          10       15   20          0     5        10       15   20
                                                delay (s)                                delay (s)
                              Evidence for delay-dependent spatial drift
                                         Participants                   Dynamic Field Model
                          10                                       10
                                center
                          8     20 degrees                         8
mean constant error (°)




                                40 degrees
                          6                                        6

                          4                                        4

                          2                                        2

                          0                                        0

                          -2                                       -2
                          7     0        5      10       15   20   7    0   5      10       15   20

                          6                                        6
mean variabl error (°)




                          5                                        5

                          4                                        4

                          3                                        3

                          2                                        2

                          1                                        1

                          0                                        0
                                0       5       10       15   20        0   5      10       15   20
                                             delay (s)                          delay (s)
                              Evidence for delay-dependent spatial drift
                                         Participants                   Dynamic Field Model
                          10                                       10
                                center
                          8     20 degrees                         8
mean constant error (°)




                                40 degrees
                          6                                        6
                                60 degrees
                          4                                        4

                          2                                        2

                          0                                        0

                          -2                                       -2
                          7     0        5      10       15   20   7    0   5      10       15   20

                          6                                        6
mean variable error (°)




                          5                                        5

                          4                                        4

                          3                                        3

                          2                                        2

                          1                                        1

                          0                                        0
                                0        5      10       15   20        0   5      10       15   20
                                             delay (s)                          delay (s)
                              Evidence for delay-dependent spatial drift
                                         Participants                   Dynamic Field Model
                          10                                       10
                                center
                          8     20 degrees                         8
mean constant error (°)




                                40 degrees
                          6                                        6
                                60 degrees
                          4     80 degrees                         4

                          2                                        2

                          0                                        0

                          -2                                       -2
                          7     0        5      10       15   20   7    0   5      10       15   20

                          6                                        6
mean variable error (°)




                          5                                        5

                          4                                        4

                          3                                        3

                          2                                        2

                          1                                        1

                          0                                        0
                                0        5      10       15   20        0   5      10       15   20
                                             delay (s)                          delay (s)
      What does stability tell us?
Example: What does zero bias and
 lower variability at 0º indicate?              10

                                                8
                                                     center
                                                     20 degrees
                                                                                            10

                                                                                            8




                      mean constant error (°)
 Hypothesis: stability created                  6

                                                4
                                                                                            6

                                                                                            4



 through real-time coupling to a                2

                                                0
                                                                                            2

                                                                                            0



 perceived reference frame                      -2
                                                7
                                                     0        5         10       15   20
                                                                                            -2
                                                                                            7
                                                                                                 0   5       10       15   20

                                                                                                         Stability at midline
                                                6                                           6

 (midline symmetry)   mean variable error (°)
                                                5

                                                4
                                                                                            5

                                                                                            4



 Recent evidence is consistent                  3

                                                2

                                                1
                                                                  Variance grows in time
                                                                  (change in stability is
                                                                                            3

                                                                                            2

                                                                                            1


 with this view                                 0
                                                     0       5
                                                                  the correct picture)
                                                                        10
                                                                     delay (s)
                                                                                 15   20
                                                                                            0
                                                                                                 0   5       10
                                                                                                          delay (s)
                                                                                                                      15   20




  – Can create and destroy effect by
    adding/removing perceptual cues in
    the task space
DFT: Reference Frames

                                                         10
                  Reference Input                        8
                                                              center
                                                              20 degrees




                              mean constant error (°)
  ∆y                                                          40 degrees
                                                         6
                                                              60 degrees
                                                         4    80 degrees
                  Reference-based
                                                         2
                     WM peak                             0

                                                         -2
                                                         7    0        5

                                                         6
                    Experiment




                               mean variable error (°)
                                                         5


                  informs theory                         4

                                                         3

                   and provides                          2

                                                         1

                    constraints                          0
                                                              0        5
                                    Is drift unique to space?
                                    Color Space
                                        0°
                                                              Target appears            Mouse Response

                                                                                          Rotated 110° arc



No! VWM for colors shows delay-dependent
 -90°          90°   -70°
                    Target
        -70°               -70° Color
drift in color space        at -35° in
                           task space
                                       +
                                        +/-180°
 Color Directional Error (º)




                                           Close Colors (N = 12)               Far Colors (N = 12)
                               5                                    5
                                                   -110º
                               3                                    3
                                                                                             -10º
                               1                                    1
                               -1                                  -1
                               -3                                  -3
                                                   -70º                                      -170º
                               -5                                  -5
                                    0         5      10      15          0         5        10       15
                                             Delay (sec)                            Delay (sec)
 Hidden Theoretical Challenge
Signatures of real-time dynamics captured by
DFT are evident along both spatial and featural
dimensions.
– Piece of cake…just re-label ―x‖, right?
Violates our commitment to embodiment
– Fails to specify how VWM remains linked to the
  sensori-motor world
    Could be a real limitation because VWM is coupled in
    real-time to perceptual structure (calibration problem)
– No account for response generation (motor system)
        Feature-Space Fields
Embodied solution: couple feature dimensions
to space via 2D feature-space fields
– Allows identification of target item in space (e.g.,
  conjunction visual search: Johnson, Spencer,
  Wilimzig, & Schöner poster)


                                            featur
                                                 e
                Are we just being anal?
   From theory to experiment: metrics in 2D matter
                            color        space

    – If estimate the same colors used previously but
                                space
      separate them spatially, should reduce drift
   Yes! Stability via coupling to space
                                                                                                     color
                                                 Figure 8. Reduced spatial errors through color

                                                                                  9




                                                   Color Directional Error (°)
      color                     space                                             8
                                                                                  7
                                                                                  6
                                                                                  5
                                                                                  4
 space                                                                            3
                                                                                  2
                                                                                  1
                                                                                  0
                                                                                 -1              -130/-110      -130/-110
                                                                                 -2    Outer        Left          Center
                                                                                      Targets   Color Target/Spatial Location

                        color                    Figure 9. Reduction of color error through space
Figure 8. Reduced spatial errors through color
Trial 7: Word-Space Binding Phase
Trial 7: Word-Space Binding Phase
Trial 7: Word-Space Binding Phase
Trial 7: Word-Space Binding Phase
Trial 7: Word-Space Binding Phase
Trial 7: Word-Space Binding Phase
Trial 8:Test Phase
Trial 8:Test Phase
Trial 8:Test Phase
Trial 8:Test Phase
Trial 8:Test Phase
Trial 8:Test Phase
Real-time Dynamics: Summary
DFT and empirical results highlight the delicate
balance between stability, instability, flexibility
– Stability: Self-sustaining peaks; coupling to
  perceptual input and sensori-motor system
– Instability: delay-dependent drift; delay-dependent
  increase in variable error
– Flexibility: use of activation dynamics to bring
  together events separated in time
Another critical stabilization process…learning
and the formation of long-term memories
Learning Dynamics

                                                          10
                                                               center
                                                          8    20 degrees




                               mean constant error (°)
∆y                                                             40 degrees
                                                          6
                                                               60 degrees
                                                          4    80 degrees

                                                          2

                                                          0

                                                          -2
                                                          7    0        5

                                                          6




                                mean variable error (°)
                                                          5


                      LTM trace of                        4


                    LTM of previous
                    ―reference‖ peak
                                                          3

                                                          2

                      ―target‖ peak                       1

                                                          0
                                                               0        5
                                               Integrating time scales
 LTM traces arise from emergent, context-
 specific ―decisions‖ (i.e., peaks)
  – Not always forming LTMs of input statistics

3 targets (20º,40º,60º)                                                                                             Spencer & Hund (2003)
Manipulated frequency




                                                                                mean constant directional difference (°)
         mean constant directional error (°)




            8                                                                                                              4
 – Freq = 2/3rds; Infreq = 1/6
            7
              a
                                                                                                                           3
                                                                                                                                b            Freq: outer




                                                                                                                                                               outward
 – Direction: Inner, Outer

                                                            away from midline
            6                                                                                                              2
            5                                                                                                              1
Center target biased in direction of
            4                                                                                                              0
frequent target, but not




                                                                                                                                                                inward
            3                                                                                                              -1
symmetrically
            2                                   BI                                                                         -2       Freq: inner
 – drift shows up in LTM traces
            1
                                                BO
                                                                                                                           -3
                inner center  outer                                                                                                  inner   center    outer
Blending of input statistics and
           -2
                       target                                                                                                                target
real-time dynamics!
  Consequences of LTM Traces
Results illustrate the delicate balance among stability,
instability, and flexibility over longer time scales
– Frequent target stabilized (low variable error too!)
– Infrequent targets show bias (high variable error)
– Blending of activation yields generalization (implicit
  categories!)
But are traces in the DFT really long-term memory?
Can they usefully guide performance in long-term
memory paradigms?
        Long-term memory
Example: cued recall
Insight(?): DFT can generate reliable
responses with graded, ―fuzzy‖ LTMs because
it has rich, real-time dynamics
                       Development
What’s the link to development?
   Schutte & Spencer (in prep)
                                     Complex pattern of
                                     reference effects over
                                     development
                                     Quantitative reduction in
                                     the range across which
Schutte, Spencer, & Schöner (2003)   we see A-not-B-type
                                     biases in sandbox
                                     Systematic decrease in
                                     VE over development
                                     Decrease in A-not-B
                                     biases in spaceship
                       Development
What’s the link to development?
   Schutte & Spencer (in prep)                    We’ve linked all of these
                                                  findings to a single
                                                  developmental hypothesis…

                                                  Spatial Precision Hypothesis
                                                 Spencer & Hund (2003);
Schutte, Spencer, & Schöner (2003)          Schutte, Spencer, & Schöner (2003)
                                                   0.25

                                                    0.2
                                                                        later development
                                                   0.15
                                     activation     0.1

                                                   0.05

                                                     0
                                                                             early development
                                                  -0.05

                                                   -0.1

                                                  -0.15

                                                   -0.2

                                                             location
        What causes the cause?
                                          Calibration/                          10

                                            alignment                           8
                                                                                     center
                                                                                     20 degrees




                                                     mean constant error (°)
             ∆y                                                                      40 degrees
                                         parameter that                         6
                                                                                     60 degrees
                                                                                4    80 degrees
                                         can transform                          2


                                            egocentric                          0

                                                                                -2
                                            reference                           7    0        5

                                                                                6
                                           frame (e.g.,




                                                      mean variable error (°)
                                                                                5


                                         retinotopic) to                        4

                                                                                3

                                         object-centered                        2

                                                                                1

                                         ref frame (e.g.,                       0
                                                                                     0        5

                                             midline)


This might be the stuff of ―non-obvious‖ causes…
The Dynamics of Reference Frame
         Alignment
                      Calibration Field =
                      activation dynamics                                                                               10

                                                                                                                        8
                                                                                                                                 center
                                                                                                                                 20 degrees

                      for Δy




                                                                                              mean constant error (°)
     ∆y                                                                                                                          40 degrees
                                                                                                                        6
                                                                                                                                 60 degrees
                                                                                                                        4        80 degrees
                                                                             10
                                                                                                                        2    center
                           Direction0(ego)
                                        20 degrees                                                      8




                                                                        mean constant error (°)
    ∆y                                                                                                                       40 degrees
                                                                                                        6
                                                                                                         -2                  60 degrees




                                                                                                                                          Direction (object)
                                                                                                        47                     0
                                                                                                                             80 degrees 5

                                                                                                        26




                                                      mean variable error (°)
                                                                                                         5
                                                                                                        0
                                                                                      4
                                                                                    -2
                                                                                    73                                       0            5

                                                                                    62




                                 mean variable error (°)
                                                                                    51

                                                                                    40
                                                                                                                                 0             5
                                                                                    3                                                                           d

                                                                                    2

                       Proposal: SPH can                                            1

                                                                                    0
                       emerge from                                                                                           0            5
                                                                                                                                                               del


                       unstable calibration
Consequences of Poor Alignment
 Reference peaks will
 build at slightly
 different places in
 the object-centered
 frame from trial to
 trial
 Schmears out
 reference-related       Target-related WM peaks become
 LTM traces              slightly mis-aligned from trial to trial
 With broad LTM,         Schmears out target-related LTM
 attraction dominates    traces which are the source of AnotB
 With narrow LTM,        Stability of calibration + sharpening
 inhibition dominates    of LTMs leads to reduction in VE and
 to left/right of axis   more robust peaks
  Constraints and Time Scales
This story reveals complex dependencies
among dynamics at different time scales…
– Real-time dynamics: calibration, WM peaks,
  ―drift‖
– Learning dynamics: broad v. narrow LTMs
– Developmental dynamics: enhanced stability of
  calibration / reference frame alignment
         What causes the cause?
Interesting, but what causes the increased stability in
  calibration dynamics over development??
  As Huttenlocher and Newcombe have noted, there are
  real consequences to mis-alignment

                          DFT integrates LTMs of
                          reference frame (context) and
                          ―items‖…alignment critical
                          for re-connecting with past
                          knowledge in a way that can
                          guide current action
       What causes the cause?
There is a rich array of feedback in daily life that can
point toward more or less stable calibration in context
given the consequences of mis-alignment, broad is
good early in dev when you are trying to figure out
what matters (less precision is more accuracy)
       Would you please get to ―real‖
           development now??
By now, our OCD is shining through…
  I suspect connectionists would’ve taken the SPH, run some
  simulations to show the emergence of narrow interaction over
  time and published!
  Moreover, I suspect some people would claim that we have yet
  to capture ―real‖ development
   – At what point should we stop to ―close the loop‖ and create systems
     that change themselves??
  That tension is good—keeps us striving for greater
  understanding; but it also fails to acknowledge what we’ve
  learned about development in the mean time!
  We’ve uncovered a rich set of empirical and theoretical
  constraints that any developmental account of these
  phenomena must explain…
          Future Challenges
Challenges for us—to see how far the rabbit
hole goes:
– What is a ―context‖ or reference frame in a multi-
  dimensional picture (beyond 1D)?
– How do we re-activate LTM in a context-specific
  manner to re-connect with our past experiences?
– How do we generalize across experiences v. form
  memories unique to the situation?
          Future Challenges
Challenges for connectionists (?):
– We assume a lot of pre-structure in our fields: do
  insights from connectionism inform us about the
  development of that structure?
– What learning/developmental processes might lead
  to context- and task-specific enhancement of our
  calibration dynamics?
          Future Challenges
Challenges for us all:
– To acknowledge our bias to search for ―primary‖
  causes of development (mechanisms?) and to
  appreciate that an understanding of development
  can come in often non-obvious ways
               Thanks to:
This research was supported by…
 NIMH RO1 MH62480 awarded to John P.
 Spencer
 NSF BCS 00-91757 awarded to John P.
 Spencer
 The Obermann Center for Advanced Studies at
 the University of Iowa

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:4
posted:11/24/2011
language:English
pages:52