Neural mechanisms of cognitive control: Convergent computational

1
Neural mechanisms of cognitive control
Convergent computational neuroimaging studies

• Todd Braver
• Cognitive Control and Psychopathology Lab,
• Washington University

Much Credit To Joshua Brown Andrew Jones Jeremy
Reynolds
2
Approach

• Understanding of high-level cognition
• attentional control, working memory, planning,
  decision making
• Focus on human behavior and brain activity
• capturing detailed performance measures
• accounting for brain imaging (e.g., fMRI) data
• Macro-scale level models
• abstract away from many potentially relevant
  biological details
• extract central computational mechanisms relevant
  to cognitive control

3
Key elements of cognitive control

• The capability to dynamically adjust to changing
  internal and external contingencies
• Internal goal representations, attentional focus
• lateral PFC -gt active maintenance of context
  information (Cohen Servan-Schreiber, 1992
  Braver, Barch Cohen, 2001 Miller Cohen,
  2001)
• Goal selection and updating
• midbrain dopamine (DA) system -gt gating signal
  (Braver Cohen, 1999 Braver Cohen, 2000
  OReilly et al., 2002)
• Performance monitoring and adjustment
• anterior cingulate cortex (ACC)-gt conflict
  detector (Carter et al., 1998 Botvinick et al.,
  2001)

4
Minimalist canonical model
Conflict Detection (ACC)
Active Memory
Control Regulation
Goal / Context (PFC)
Response (Motor Ctx)
Bias
Selection/ Updating
Association (Post. Ctx)
Learning
Input
Reward
5
Performance Monitoring Adjustment

• A central element of controlled intelligent
  behavior is the ability to self-monitor and
  dynamically regulate performance
• Changing strategies
• Preventing potential mistakes
• Reducing tentativeness
• Examples
• Reacting to opponents opening gambits in
  chess-playing
• Driving on icy roads
• What are the computational and neural mechanisms
  by which humans monitor and adjust performance
  on-line?

6
Conflict and Performance Monitoring

• What information can be used to monitor
  performance?
• error or reinforcement
• however, not always available
• require internal evaluation measure
• Conflict as performance evaluation measure
• old idea from information-theory (Berlyne, 1960)
• quantify conflict as Hopfield energy in response
  system
• -????ai aj wij
• Hypothesis
• conflict is computed in ACC
• conflict information can be used by other systems
  to adjust control strategies

7
From Conflict to Control

• Feedback loop
• environmental conditions elicit conflict
• conflict detection (ACC) recruits increased
  control (PFC)
• increased control reduces experienced conflict

8
ACC function

• Neuroimaging data suggests role in cognitive
  control
• activation related to task difficulty
• occurs in wide variety of task domains
• attention, memory, language, learning, etc.
• activation data consistent with conflict
  hypothesis (e.g. Stroop)
• however, ACC probably involved in other functions
  besides conflict (e.g., emotional processing)

9
Sequential Choice Responding

• Goals
• Identify significant trial-by-trial fluctuations
  in conflict during cognitive task performance
• Demonstrate that conflict fluctuations are
  reflected in terms of ACC activity fluctuations
• Demonstrate that conflict fluctuations are
  associated with adjustments in control strategy
  and performance
• Task
• sequential reaction time task
• two-alternative forced-choice (TAFC)
  discrimination
• Rationale
• simple task, not easy from higher cognition
  standpoint, but easy to simulate
• provides boundary conditions for more complex
  models

10
Performance Data - Global
Reaction Time
Error Rate
Performance as a function of global stimulus
probability High (51 ratio), Equal (11 ratio),
Low (15 ratio)
11
ACC Activity -- Global Frequency
Braver et al., 2001 Cerebral Cortex
12
Performance Data - Local Sequence
Reaction Time
Errors
Performance as a function of local stimulus
sequence 5-trial history (RRepeat, AAlternate)
Jones, Cho, Nystrom, Cohen and Braver Cognitive,
Affective and Behavioral Neuroscience, in press
13
Task Model
14
Priming Conflict Mechanisms

• Unit activation function (Usher McClelland,
  2001)
• Xt(n) Xt(n-1) ? ? -( k ? Xt(n-1) ) (? ?
  f( Yt(n-1) ) ) ?t(n) It(n) St BtX
• f(x) linear activation function k leak
  parameter ? lateral inhibition strength
• It(n) preceding layer input ?t(n) noise
  parameter St and Bt strategic sequential
  priming mechanisms
• Sequential Priming (Cho et al., in press)
• BtX RtX At RtX ? ? R(t-1)X (1-?) ? MR ?
  ?(t-1)X At ? ? A(t-1) (1-? ) ? MA ? ?(t-1)
• Rtx repetition priming At alternation priming
  ? time constant
• MR, MA maximum repetition
  and alternation priming
• ?tX repetition detector ?t alternation
  detector
• (step function 1 if response(t)X, 0 if not X)
  (step function 1 if response(t) alternation,
  0 if not)
• Conflict
• Et ?n X(n) ? Y(n)
• Strategic Priming (control adjustment) (Botvinick
  et al., 2001)
• St ? ? S(t-1) (1-?) ? ? E(t-1) ?

15
Capturing Sequential Behavior
Reaction Time
Error Rate
R2 0.87
R2 0.79
Jones et al., CABN in press
16
ACC Response -- Global Frequency
Imaging Data
Simulation Data
17
Capturing ACC Activity
R2 0.73
Jones et al., CABN in press
18
Capturing control adjustment effects
Jones et al., CABN in press
19
Cognitive Control Task-switching

• Rapid switching between tasks that share
  perceptual and response characteristics occurs
  often in everyday life
• Example writing programming code and writing
  email
• Explosion of recent studies in cognitive
  psychology cognitive neuroscience
• Task-switching thought to involve cognitive
  control
• Switch costs preparatory interval
• Internal representation and reconfiguration of
  task-sets
• Goals
• Test for commonalities between task-switching and
  TAFC
• Conflict-control loops related to speed-accuracy
  shifts -gt switching alternation
• Test for differences between task-switching and
  TAFC
• Task-set priming -gt incongruency effects
  (activation of irrelevant task pathway)
• Control adjustments related to attentional shifts
  -gt locking in task-sets
• Hypotheses
• Dual conflict detectors in ACC
• Response conflict, task-set conflict
• Different mechanisms for control adjustment
• Speed-accuracy shifts, attentional sharpening

20
Task Switching Paradigm
TASK A
TASK B
LETTER
NUMBER
X 9
X 9
Time
Time
Vowel
Cons
Odd
Even
21
Sequential Performance Data
Effects of Previous Trial Type on Current Trial
Performance
Congruency Effects
Alternation Effects
Switching Effects
22
Model
Output ACC (ACCo)
Task Set ACC (ACCt)
Remember to describe development of RT
framework, parallel modelling streams
23
Modeling Results

• Attempted to fit 64 data points
• 3 factors
• Switch/NoSwitch
• Repeat/Alternate
• Congruent/Incongruent
• Current trial as a function of previous trial
• R2 .73
• Within 95 confidence intervals for 61/ 64 data
  points
• Max z-score of model vs. data 2.68

Model RT
Behavioral RT
24
Capturing Performance Data
25
Capturing Sequential Effects
Model data under intact and ACC- lesion conditions
Switch Costs - Incongruent Trials
Alternation Costs
Switch Costs
26
ACC Activity fMRI Predictions
27
Summary

• ACC involved in conflict detection
• Model fits detailed aspects of TAFC behavioral
  data
• Model predicts ACC modulation related to
  trial-by-trial fluctuations in conflict
• Task-switching and cognitive control
• Conflict might involve both competing responses
  and competing task-sets
• Dual mechanisms of control in task-switching
• Speed-accuracy shifts, attentional focus
• Can account for sequential effects in behavioral
  performance
• Makes predictions regarding brain activation (to
  be tested)
• Simple computational framework can provide
  guidance regarding neural mechanisms of cognitive
  control
• Can be used for hypothesis generation and testing
  in neuroimaging experiments