N-Smith-Sloan-Swartz-Annual-2008

Simultaneous integration
versus sequential sampling in
multiple-choice decision
making
Nate Smith
July 20, 2008
Decision making
• A cognitive process of choosing an
opinion or action between ≥2 choices
• Simultaneous integration accumulates
evidence for both choices
• Sequential sampling dependent upon
active changes in attention for choice
action
Decision making
Simultaneous integration
Decision making
Sequential Sampling
Decision making
Sequential Sampling
Decision making
Sequential Sampling
Decision making
Sequential Sampling
Simultaneous Integration
Accumulator models used in
perceptual decision making
Diffusion Model
•Does not easily extend to N-choice
Leaky Competing Accumulator Model
•Does not retain ‘early’ information
•Can a network of neurons produce N-choice behavior?
Smith and Ratcliffe, 2004
Reduced 2 variable model
for perceptual discrimination
Mean
field
approx.
Simplified F-I curves
Slow NMDA
gating variable
Constant NS activity
Reduced two variable model
Wong and Wang, 2006
Generalized N-choice model
for perceptual decisions
Multiple alternative simultaneous
integration decision making
•Similar to previous
random-dot motion
tasks
•Three directions of
coherent motion
•Subject has to
saccade in direction
of highest perceived
motion (highest
coherence)
Niwa and Ditterich, 2008
Performance dependent on
overall motion
Niwa and Ditterich, 2008
•Psychometric and reaction time data are more complex
•Simpler mechanism for describing choice behavior?
Research aims
• Can a biophysically realistic neural
mechanism reproduce results similar to
the human psychophysics study?
•Investigate whether the psychometric
softmax function holds for N-choice tasks
P a  
1
N
1  e

n1
•What dynamics underlie N-choice
decision making?
(ca  cn )

Neural data produces variable
reaction times and decisions
3-choice model fits human
psychophysics data
•Neural model is
able to reproduce
findings from 3choice
simultaneous
integration task
Theoretical psychometric
softmax function fits data
P a  
1
N
1  e

(ca  cn )

n1
•Plotting for different coherence values matches up vs.
softmax function
Reaction time data
 0  35
0  25  0.375(max(coh)  min(coh))
Possible lateral inhibition/modulation in area MT responsible
for scaling of input with multiple signals?
Sequential Sampling
Neural activity integrates
information from each gaze
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
Neural activity integrates
information from each gaze
A
B
First gaze biases selection
and reaction time
•First gaze increases chance of choosing an option when
objects have equivalent value
Mean reaction time (ms)
Probability
•Reaction time for objects with first gaze faster
Conclusions
• Biophysically realistic reduced model
replicates experimental data
• Softmax function can work as a general
underlying framework for decision
making in neural circuits
• Neural pools can retain and integrate
information even in absence of fixation
Acknowledgments
Wang Lab
Xiao-Jing Wang
Alberto Bernacchia
Tatiana Engel
Morrie Furman
John Murray
Chung-Chuan Lo
Christian Luhmann
Jacinto Pereira
Dahui Wang