MEGaVis: Perceptual Decisions in the Face of Explicit Costs and Benefits Michael S. Landy Julia Trommershäuser Laurence T. Maloney Ross Goutcher Pascal Mamassian
Statistical/Optimal Models in Vision & Action Sequential Ideal Observer Analysis Statistical Models of Cue Combination Statistical Models of Movement Planning and Control –Minimum variance movement planning/control –MEGaMove – Maximum Expected Gain model for Movement planning
Statistical/Optimal Models in Vision & Action MEGaMove – Maximum Expected Gain model for Movement planning –A choice of movement plan fixes the probabilities p i of each possible outcome i with gain G i –The resulting expected gain EG=p 1 G 1 +p 2 G 2 +… –A movement plan is chosen to maximize EG –Uncertainty of outcome is due to both perceptual and motor variability –Subjects are typically optimal for pointing tasks
Statistical/Optimal Models in Vision & Action MEGaMove – Maximum Expected Gain model for Movement planning MEGaVis – Maximum Expected Gain model for Visual estimation –Task: Orientation estimation, method of adjustment –Do subjects remain optimal when motor variability is minimized? –Do subjects remain optimal when visual reliability is manipulated?
Task – Orientation Estimation
Payoff (100 points) Penalty (0, -100 or -500 points, in separate blocks)
Task – Orientation Estimation Payoff (100 points) Penalty (0, -100 or -500 points, in separate blocks)
Task – Orientation Estimation
Done!
Task – Orientation Estimation
100
Task – Orientation Estimation -400
Task – Orientation Estimation -500
Task – Orientation Estimation Align the white arcs with the remembered mean orientation to earn points Avoid alignment with the black arcs to avoid the penalty Feedback provided as to whether the payoff, penalty, both or neither were awarded
Task – Orientation Estimation Three levels of orientation variability –Von Mises κ values of 500, 50 and 5 –Corresponding standard deviations of 2.6, 8 and 27 deg Two spatial configurations of white target arc and black penalty arc (abutting or half overlapped) Three penalty levels: 0, 100 and 500 points One payoff level: 100 points
Stimulus – Orientation Variability κ = 500, σ = 2.6 deg
Stimulus – Orientation Variability κ = 50, σ = 8 deg
Stimulus – Orientation Variability κ = 5, σ = 27 deg
Payoff/Penalty Configurations
Where should you “aim”? Penalty = 0 case Payoff (100 points) Penalty (0 points)
Where should you “aim”? Penalty = -100 case Payoff (100 points) Penalty (-100 points)
Where should you “aim”? Penalty = -500 case Payoff (100 points) Penalty (-500 points)
Where should you “aim”? Penalty = -500, overlapped penalty case Payoff (100 points) Penalty (-500 points)
Where should you “aim”? Penalty = -500, overlapped penalty, high image noise case Payoff (100 points) Penalty (-500 points)
Experiment 1 – Variability
Experiment 1 – Setting Shifts (HB)
Experiment 1 – Score (HB)
Experiment 1 – Setting Shifts (MSL)
Experiment 1 – Score (MSL)
Experiment 1 – Setting Shifts (3 more subjects)
Experiment 1 – Score (3 more subjects)
Experiment 1 - Efficiency
Intermediate Conclusions Subjects are by and large near-optimal in this task That means they take into account their own variability in each condition as well as the penalty level and payoff/penalty configuration Can they respond to changing variability on a trial-by-trial basis? → Re-run using a mixed-list design (all noise levels mixed together in a block; only penalty level is blocked)
Experiment 2 – Setting Shifts (HB)
Experiment 2 – Score (HB)
Experiment 2 – Setting Shifts (MSL)
Experiment 2 – Score (MSL)
Experiment 2 – Setting Shifts (2 more subjects)
Experiment 2 – Score (2 more subjects)
Experiment 2 - Efficiency
Conclusion Subjects are nearly optimal in all conditions Thus, effectively they are able to calculate and maximize effective gain across a variety of target/penalty configurations, penalty values and stimulus uncertainties The main sub-optimality is an unwillingness to “aim” outside of the target This is “risk-seeking” behavior, unlike what is seen in paper-and-pencil decision tasks