D ECIDING WHEN TO CUT YOUR LOSSES Matt Cieslak, Tobias Kluth, Maren Stiels & Daniel Wood
O UTLINE I. Introduction II. Model III. Experiment IV. Results V. Conclusion 2
R ESEARCH Q UESTIONS 1. Are people optimal when they decide to cut their losses? 2. Does the GSR influence the optimality? 3 ! ! ! ? ?
D ECISION M AKING M ODELS Classic “diffusion” model Accumulate all evidence: Compare to a constant threshold / accuracy criterion Urgency Gating model Accumulate only the novel evidence: Compare to a dropping accuracy criterion 4
U RGENCY G ATING M ODEL Compute estimate of evidence summation (≈ integration!) of new information low-pass filtering (to deal with noise) “temporal filter model” (Ludwig et al J. Neurosci 25: ) Multiply by growing function of time and compare to a threshold 5
S ETUP 6 4 ½ feet GSR2* 13‘‘ at 30 Hz * GSR2: Device to measure the galvanic skin resonse and sampled at 44.1 kHz Response by the keyboard with the buttons ⟵ and ⟶ 7 subjects
D ESIGN 7 End of trial by response or Time-out after 5 sec or 8 sec Time Duration of a trial 5 or 8 sec Random uniform distribution was used for the onset of dots Dots were presented on 60% of the trials Duration (random): 1-5 sec (Dot-trial) 5 or 8 sec (Time-out-trial)
C ONNECTING TO THE U RGENCY -G ATING M ODEL Time out -35 Points t=0t=t t=t end t end E(t) t end E(t) t end E(t) dots no dots dots no dots dots no dots
C ONNECTING TO THE U RGENCY -G ATING M ODEL Correct 20 Points t=0t=t t=t end t end E(t) t end E(t) t end E(t) dots no dots dots no dots dots no dots u(t)
10 t end E(t)
11 8 sec 5 sec Trial length
R ESULTS GSR predicted the latency of their guess on no-dot trials Response-time decreased linearly by a function of time 12
C ONCLUSION 13 2 types of subjects: Just guess:uncertainly not handled well or time feeling very bad Wait: good estimate of time; optimal behaviour High GSR does not predict an early response Instead it appears to increase as the person waits Provides evidence for an urgency signal
L ITERATURE Lecture Slides ‚The blurry borders between decision and doing‘ (Part I, Part II) of Paul Cisek at the CoSMo Summer School 2011 Cisek, Puskas and El-Murr Pictures road/ thinker1.jpeg net/quatsch/bilder/computer/windows_wait.jpg 14
H IGHSCORE 15 Thank you! # subject5 sec Version # subjects8 sec Version
C LASSIC M ODELS Well-supported by data like behavioral data (error rates, reaction time distributions) neural activity Similar to the sequential probability ratio test (SPRT) optimal for requiring the fewest samples to reach a given criterion of accuracy Widely accepted conclusion: “Diffusion model explains decisions” 16
S UMMARY Serial model: When Cognition is done, action can begin i.e. “decision threshold” But what controls growth toward the threshold is an urgency signal i.e. a signal related to motor initiation When reaching a motor initiation threshold, we commit to our current best guess Cognition and Action are not so separate 17
U RGENCY G ATING M ODEL Addition of a criterion of confidence that drops over time Results confirm urgency-gating model over integrator models Cisek, Puskas and El-Murr, 2009 Previous results with constant-evidence tasks compatible with both models Error rates Reaction time distributions Neural activity in LIP, SC, PFC, etc. Optimization of reward rate, and redundancy between samples Proposed to be responsible for observed neural activity growth/distributions of RTs 18