10 / 31 Outline Perception workshop groups Signal detection theory Scheduling meetings
Detection experiment Question –How sensitive is an observer to a sensory stimulus; for example, light?
Detection experiment Question –How sensitive is an observer to (for example) light? Classic experiment –Yes/No task
Detection experiment Question –How sensitive is an observer to (for example) light? Classic experiment –Yes/No task –Measure threshold intensity needed to have 50% hits
Threshold
Jane Nancy
Summary of results Thresholds –Jane = 20 –Nancy = 25
Summary of results Thresholds –Jane = 20 –Nancy = 25 False alarm rates –Jane = 51% –Nancy = 18.7%
Look at one intensity level I = 25
Jane’s Hit Rate P(H) =.84
Nancy’s Hit Rate P(H) =.5
Look at one intensity level I = 25 –Jane Hit rate: P(H) =.84
Look at one intensity level I = 25 –Jane Hit rate: P(H) =.84 False alarm rate: P(FA) =.51
Look at one intensity level I = 25 –Jane Hit rate: P(H) =.84 False alarm rate: P(FA) =.51 –Nancy Hit rate: P(H) =.5
Look at one intensity level I = 25 –Jane Hit rate: P(H) =.84 False alarm rate: P(FA) =.51 –Nancy Hit rate: P(H) =.5 False alarm rate: P(FA) =.187
Signal detection theory terms Hits - p(H) –Proportion of “yes” responses when signal is present
Signal detection theory terms Hits - p(H) –Proportion of “yes” responses when signal is present Misses - p(M) –Proportion of “no” responses when signal is present
Signal detection theory terms Hits - p(H) –Proportion of “yes” responses when signal is present Misses - p(M) –Proportion of “no” responses when signal is present False alarms - p(FA) –Proportion of “yes” responses when signal is not present
Signal detection theory terms Hits - p(H) –Proportion of “yes” responses when signal is present Misses - p(M) –Proportion of “no” responses when signal is present False alarms - p(FA) –Proportion of “yes” responses when signal is not present Correct rejections - p(CR) –Proportion of “no” responses when signal is not present
Relationships between terms P(H) + P(M) = 1
Relationships between terms P(H) + P(M) = 1 P(FA) + P(CR) = 1
Relationships between terms P(H) + P(M) = 1 P(FA) + P(CR) = 1 Only need to specify P(H) and P(FA)
Extreme detection strategies Most liberal (always say yes)
Extreme detection strategies Most liberal (always say yes) –P(H) = 1, P(FA) = 1
Extreme detection strategies Most liberal (always say yes) –P(H) = 1, P(FA) = 1 Most conservative (always say no)
Extreme detection strategies Most liberal (always say yes) –P(H) = 1, P(FA) = 1 Most conservative (always say no) –P(H) = 0, P(FA) = 0
Signal Detection Theory
Assume an internal measure of signal strength.
Signal Detection Theory Assume an internal measure of signal strength (X). –E.g. firing rate of ganglion cell
Signal Detection Theory Assume an internal measure of signal strength (X). –E.g. firing rate of ganglion cell X is corrupted by noise
Signal Detection Theory Assume an internal measure of signal strength (X). –E.g. firing rate of ganglion cell X is corrupted by noise –E.g. random variations in firing rate
Signal Detection Theory Assume an internal measure of signal strength (X). –E.g. firing rate of ganglion cell X is corrupted by noise –E.g. random variations in firing rate When signal is not present, X = X 0 + N
Signal Detection Theory Assume an internal measure of signal strength (X). –E.g. firing rate of ganglion cell X is corrupted by noise –E.g. random variations in firing rate When signal is not present, X = X 0 + N When signal is present, X = X S + N
o Firing rate when signal is present o Firing rate when signal is not present
Criterion Set a criterion level, C
Criterion Set a criterion level, C If X > C –Report a signal
Criterion Set a criterion level, C If X > C –Report a signal If X < C –Report no signal
o Firing rate when signal is present o Firing rate when signal is not present C=20, Liberal criterion
Liberal criterion = High hit rate
Liberal criterion = High false alarm rate
o Firing rate when signal is present o Firing rate when signal is not present C=30, Conservative criterion
Conservative criterion = Low hit rate
Conservative criterion = Low false alarm rate
Probability distribution on X (no signal)
Probability distribution on X (signal)
Liberal criterion
Conservative criterion
ROC curve
A B C No signal Signal
A B C
A B C
A B C No signal Signal
A B C
A B C
Determinants of performance No signal Signal
Determinants of performance XNXN XSXS No signal Signal
Determinants of performance XNXN XSXS ∆X No signal Signal
Determinants of performance XNXN XSXS 1. Difference in average strength of Signal measure ∆X = X S - X N ∆X No signal Signal
Determinants of performance 1. Difference in average strength of Signal measure ∆X = X S - X N 2. Amount of noise ∆X No signal Signal
Determinants of performance 1. Difference in average strength of Signal measure ∆X = X S - X N 2. Amount of noise 3. Sensitivity d’ = ∆X / ∆X No signal Signal
D’ determines which ROC curve your data will fall on
d’ =.83 d’ = 1.2 d’ = 2.5 D’ determines which ROC curve your data will fall on
Criterion determines where your data will sit on an ROC curve
Conservative criterion Liberal criterion Criterion determines where your data will sit on an ROC curve
Measuring sensitivity
Pick a stimulus level for a yes / no task
Measuring sensitivity Pick a stimulus level for a yes / no task Measure hit rate and false alarm rate
Measuring sensitivity Pick a stimulus level for a yes / no task Measure hit rate and false alarm rate Use p(H) and p(FA) to calculate d’
Measuring sensitivity Pick a stimulus level for a yes / no task Measure hit rate and false alarm rate Use p(H) and p(FA) to calculate d’ d’ = absolute measure of sensitivity
Blood test example Get a blood test for level of protein A.
Blood test example Get a blood test for level of protein A. Doctor says that test is positive for liver cancer.
Blood test example Get a blood test for level of protein A. Doctor says that test is positive for liver cancer. Doctor recommends surgery to collect tissue sample for biopsy.
Blood test example Get a blood test for level of protein A. Doctor says that test is positive for liver cancer. Doctor recommends surgery to collect tissue sample for biopsy. What should you ask the doctor about the blood test?
No cancer Cancer
Liberal criterion No cancer Cancer
Conservative criterion No cancer Cancer