Modeling the Modelfest Data

1
Modeling the Modelfest Data

• Decoupled Probability Summation
• Laura L. Walker, Stanley A. Klein, UC Berkeley
• Thom Carney, Neurometrics Institute UC Berkeley

2
GOALS

• Develop a robust model of human visual detection
• Utilize the Modelfest database
• Determine optimal filter and probability
  summation
• Characterize parameters
• Filter aspect ratio (slength, swidth)
• Spatial pooling (ps)
• Mechanism pooling (pm)

3
SAMPLE OF 43 STIMULI
4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
FIT ERROR OPTIMAL FILTER BANK
MODEL
t-test values
HUMAN
Stimulus
8
(No Transcript)
9
(No Transcript)
10
OPTIMAL FILTER POOLING
11
(No Transcript)
12
CONCLUSIONS

• Standard filter model makes accurate preditions
• Few inaccurate predictions may be caused by model
  limitations or data collection limitations
• Need stimuli that will constrain slength

13
MODELFEST INFO

• Session SuE
• 100pm - 300pm
• Room 203/204
• http//www.neurometrics.com/

14
FIT ERROR
ps 3.2
t-test values
Stimulus
15
FIT ERROR OPTIMAL FILTER
t-test values
Stimulus
16
LENGTH TUNING STIMULI
DFit Error
Stimulus
17
FIT ERROR
ps 2.0
t-test values
Stimulus
18
FIT ERROR
ps 4.0
ps 3.2
ps 2.0
t-test values
Stimulus