Robust averaging during perceptual choice Vincent de Gardelle and Christopher Summerfield
[email protected] http://sites.google.com/site/vincentdegardelle/
Robust averaging
Background: SDT and diffusion
using multiple logistic regression we assessed the weight of evidence for each element, as a function of its rank in the array.
Signal detection theory (SDT) as a model for perceptual choices
Accumulation models using SDT in time
LPR model
Robust averaging: variance effect
Questions: how do we integrate multiple elements? process the mean and variance of the evidence?
within participants
across participants
1 stimulus rank
1.5
Multi-element averaging paradigm RT(s)
beta reduction
beta
inlying outlying
0.5
1
0.5
0
Errors(%) −0.5
decision relevant
30
0.7
15 0.6 0
fixation
low mean |μ|=1 high mean |μ|=2
stimulus
30
0.7 feedback
15 0.6
σ x>0
low mid high variance
0
low
med
0
high
bars: human data, lines: LPR model
40
80
120
160
variance cost on RT (ms)
variance
dots: individual subjects, line: regression
Robust averaging: within trial 1. we computed separately the weighting profiles for the two categories of trials... 1.2
within-trial discarding: separate profiles
1
low mid high variance
0.8
beta
μ x<0
decision irrelevant
response is the average more red or blue?
0
2. Control experiment: testing different regions of the color space, discarding was related to the decision value, not to the color itself.
0.6 0.4 0.2
μ<0 μ>0
0 −0.2
DV(t+1) = DV(t) + inc + noise predicted
simple average inc = μ
signal over noise inc = μ/σ
probabilistic averaging inc = mean(LPR(x))
fitted
−0.2
0
0.2
value space
between-trials discarding: global profile across categories
Conlusions When combining several pieces of evidence, subjects exhibit a robust averaging behaviour, and discard the elements that contribute extreme values in the pool. In the present study, this behaviour can be captured by a model accumulating probabilistic information across the array to make its choice.
de Gardelle, V., and Summerfield, C., (2011) Robust averaging during perceptual judgment, PNAS 108 (32) 13341-13346