信号検出理論の解説 (signal detection theory, a primer)
DESCRIPTION
駒場集中講義用資料。1/10 第2回講義分です。英語です。古いバージョンは消さずにこちらに誘導。TRANSCRIPT
SDT primer
You have a sensor (1D continuous value).You have to decide which is a signal and which
is a noise, based on the sensor value.When you classify the data as signal,
you are aware of the signal.
SDT primer
You have a sensor (1D continuous value).You have to decide which is a signal and which
is a noise, based on the sensor value.When you classify the data as signal,
you are aware of the signal.
1) You collect samples.2) You set the criteria for optimal discrimination.3) You classify new data by comparing the
sensor value and the criteria.
1) You collect samples.
1) You collect samples.2) You set the criteria for optimal discrimination.
1) You collect samples (training data).2) You set the criteria for optimal discrimination.3) You classify new data by comparing the sensor value and
the criteria.
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
SDT primer
1) You collect samples (training data).2) You set the criteria for optimal discrimination.3) You classify test data by comparing the sensor value and
the criteria.
When you classify the data as signal,you are aware of the signal.
Let’s do it again, with different data set.
1) You collect samples (training data).
1) You collect samples (training data).2) You set the criteria for optimal discrimination.
1) You collect samples.2) You set the criteria for optimal discrimination.3) You classify new data by comparing the sensor value and
the criteria.
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
3) You classify new data by comparing the sensor value and the criteria.
Criteria
Unknown processes
Recognition model (=> model-free)
Data (signal or noise)
classifywith criteria (c=2)awareness as decision
generate
Processes with unknown parametersNoise: N(0,1); Signal: N(d’,1)
Generative model (=> model-based)
Data (signal or noise)
Estimate parameter(d’ = 4) andclassify with criteria
generate
SDT primer
Processes with unknown parametersNoise: N(0,1); Signal: N(d’,1)
SDT primer
The sensitivity of the sensor is characterized as d’.
d’ is independent of criteria (c).(The correct ratio depends on c.)
Processes with unknown parametersNoise: N(0,1); Signal: N(d’,1)
SDT primer
The sensitivity of the sensor is characterized as d’.
d’ is independent of criteria (c).(The correct ratio depends on c.)
OK, but we have no such sensor.How to estimate d’ in psychophysics?
Processes with unknown parametersNoise: N(0,1); Signal: N(d’,1)
By changing criteria
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).3) You obtain data set 1 (with hit, miss, FA, CR).4) Repeat 1)-3) with different criteria.5) You reconstruct the distribution of samples.6) You estimate d’.
By changing criteria
1) You set a criterion and classify samples.
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
signal noise
yes hit ● FA ○no miss ○ CR ●
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).
1) You set a criterion and classify samples.2) You get the feedback (correct or incorrect).3) You obtain data set 1 (with hit, miss, FA, CR).
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
4) Repeat 1)-3) with different criteria.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
5) You reconstruct the distribution of samples.6) You estimate d’.
How do you change the criteria?
How do you change the criteria?1) Confidence rating (Human study)
How do you change the criteria?1) Confidence rating (Human study)
How do you change the criteria?1) Confidence rating (Human study)
YesNo
How do you change the criteria?1) Confidence rating (Human study)
Very sureUncertain SureVery sure UncertainSure
How do you change the criteria?1) Confidence rating (Human study)
2) By changing value or probability (animal study)
Very sureUncertain SureVery sure UncertainSure
How do you change the criteria?1) Confidence rating (Human study)
2) By changing value or probability (animal study)
Very sureUncertain SureVery sure UncertainSure
How do you change the criteria?1) Confidence rating (Human study)
2) By changing value or probability (animal study)
Very sureUncertain SureVery sure UncertainSure
How do you change the criteria?1) Confidence rating (Human study)
2) By changing value or probability (animal study)
Very sureUncertain SureVery sure UncertainSure
How do you change the criteria?1) Confidence rating (Human study)
2) By changing value or probability (animal study)
Very sureUncertain SureVery sure UncertainSure