Download presentation
Presentation is loading. Please wait.
Published byWidya Budiaman Modified over 6 years ago
1
Using Time-Varying Motion Stimuli to Explore Decision Dynamics
Marius Usher, Juan Gao, Rebecca Tortell, and James L. McClelland
2
Time-accuracy curves in the time-controlled paradigm
Easy Medium Hard Curve for each condition is well fit by a shifted exponential approach to asymptote: d’(t) = d’asy(1-e-(t-T0)/t)
3
Usher and McClelland (2001) Leaky Competing Accumulator Model
Inspired by known neural mechanisms Addresses the process of deciding between two alternatives based on external input (r1 + r2 = 1) with leakage, mutual inhibition, and noise: dx1/dt = r1-k(x1)–bf(x2)+x1 dx2/dt = r2-k(x2)–bf(x1)+x2 f(x) = [x]+ r1 r2 X1 X2
4
Leak and Inhibition Dominant LCA: Both can fit the d’ data
Participant chooses the most active accumulator when the go cue occurs This is equivalent to choosing response 1 iff x1-x2 > 0 Non-linearity at 0 is neglected for analytic tractability Graphs track this difference variable for a single difficulty level when the motion is to the left (Red) or to the right (Blue) d’(t) = (m1(t) – m2(t))/s(t); s(0) > 0
5
Kiani, Tanks and Shadlen 2008
Random motion stimuli of different coherences. Stimulus duration follows an exponential distribution. ‘go’ cue can occur at stimulus offset; response must occur within 500 msec to ear reward.
6
The earlier the pulse, the more it matters (Kiani et al, 2008)
7
These results rule out leak dominance
Still viable X
8
Our Preferred Model: Non-Linear LCA , with Inhibition > Leak
Final time slice
9
However, there is another interpretation
x t > Bounded Integration (Ratcliff 1999; Kiani et.al.2008)
10
Our Questions Can we distinguish the models?
Can we push around the effect?
11
Our Experiments Repeat Kiani 2008 with human subjects.
The effect was small... Let’s try a stronger manipulation. Now we have a big effect: Can we reverse or eliminate it?
12
Ongoing Investigations
Random dot motion stimuli, like those used by Shadlen and Newsome, Kiani et al, and many others. Multiple coherences: 6.4%, 12.8%, 25.6%, 51.2% Three participants per experiment, each run for up to 25 sessions. Data shown are after performance stabilizes, after varying numbers of sessions. Ongoing recruitment, Ongoing analysis…
13
Kiani Replication Exponential distribution of trial durations
Go cue when motion stops Participant must response within 300 msec of go cue and must be correct to earn a point Pulse occurs on a subset of trials, at a random time within the trial: Motion increment of +/-2% for 200 msec.
14
Our Best Participant mt
15
Experiment 2: A Stronger Manipulation
Stimulus Duration Three motion conditions crossed with 8 coherences. LCALD and BI both predict Early > Late Data shown are percent correct, averaged across coherences We include a switch condition with 6.4% and 12.8% coherences only (no right answer). LCALD and BI both predict %Early Choices > 50% Each participant has at least 600 trials per data point over at least 10 sessions. 1) Early 2) Late 3) Constant 4) Switch
16
Results in Exp.2: Star Subject
MT 1200 trials per data point. 20 sessions.
17
Results in Exp.2: Star Subject
MT
18
Results in Exp.2 CS 600 trials per data point. 10 sessions
19
Results in Exp.2 CS
20
Results in Exp.2 SC 600 trials per data point. 10 sessions.
21
Results in Exp.2 SC 600 trials per data point. 10 sessions. 21
22
Take home message Yes, it seems earlier > later in all three subjects with this time pressure. But 2 of 3 participants show some sensitivity to late information even at longer durations, while one does not. Model accounts for individual differences: BI: Low vs. high bound LCALD: strong vs. weak inhibition dominance
23
Our Experiments Repeat Kiani 2008 with human subjects.
The effect was small... Let’s try a stronger manipulation. Now we have a big effect: Can we reverse or eliminate it?
24
Experiment 3: Time-limited integration without time pressure to respond
Same stimulus conditions as before. New participants. Only two procedural changes: Uniform vs. exponential distribution of stimulus durations Participants have a full second after the end of the stimulus to respond.
25
Results in Exp.3, without time pressure
MM
26
Results in Exp.3, without time pressure
MM 25 session, 1500 trials per point
27
Results in Exp.3, without time pressure
WW
28
Results in Exp.3, without time pressure
WW 10 sessions.
29
Results in Exp.3, without time pressure
DG 15 sessions, 900 trials
30
Results in Exp.3, without time pressure
DG 15 sessions, 900 trials
31
Our Questions Can we distinguish the models?
Not yet Can we push around the effect? Yes
32
How do the models explain the data?
BI: participants can perform unbounded integration if there is no time pressure LCALD: participants can balance leak and inhibition if there is no time pressure In both cases, it appears that we have balanced, unbounded integration.
33
Two remaining questions
Can we create a situation in which we will observe leaky integration? Very long trials? Detect motion pulse in otherwise 0% background? Why does accuracy level off with long integration times if there is perfect integration? Between trial drift variance?? (Ratcliff, 1978).
34
The Bottom Line The dynamics of information integration might not be fixed characteristics of the decision making mechanism Instead, they may be tunable in response to task demands: Leak vs. competition Presence of a bound on integration Etc.
35
r1 r2 X1 X2 The End
38
Results in Exp 1. The pulse study
SC
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.