The flashcards below were created by user
on FreezingBlue Flashcards.
When a target is located it becomes necessary to detect (confirm) whether the target is an item of interest (signal) or not (noise)
the relevant, important information
all stimuli, whether relevant or not
Detect signal from noise (using his/her visual, auditory or tactile skills)
Uncertainty in whether signal is present is cause by?
- Noise in the world (bottom-up processing)
- Prior knowledge, expectancies, biases (top-down processing)
Signal Detection Theory
Model of how humans separate "noise" from "signal". Humans collect observation (using senses) then determine if there is enough evidence to constitute a signal.
- Safety Inspection: searching a crack in airplane
- Medical diagnosis
- Airport baggage check
- Manufacturing products that meet company requirements
- Photo processing (red eye)
- Medical diagnosis: fever
- Vibration sensors: vibration of car wheel
- Is something burning
- Gas leak
- Something is rotten
Assumptions in SDT
- probability of observing signal or noise is normally distributed, random noise
- signal increases the mean of the distribution
2 x 2 matrix, explain hit, false alarm, correct rejection, miss
- hit: detect / is signal
- miss: dont detect / is signal
- false alarm: detect / not signal
- correct rejection: dont detect / not signal
- yes: hit or false alarm
- no: correct rejection or miss
- how good an operator is at separating signal from noise (bottom up)
- d' = z(H) - z(FA)
What influences sensitivity (signal detection performance)
- sound to noise ratio
- redundancy of signal
- operator alertness and motivation
- cost of false alarms or misses
want to reduce overlap in the distribution of signal and noise?
- increase intensity of signal (better and stronger evidence in a criminal case)
- Improve ability of observer (training)
Response Criteria, Bias (beta)
Bias of the operator to respond "yes" or "no" to signal due to expectancy, payoffs, incentives, experiences (top-down)
Beta = Ordinary (Z(P(hit)) / Ordinary (Z(P(FA)))
Liberal vs Conservative
- Liberal - prone to say signal (risky) tendency to left
- Conservative - not prone to say signal (safe) tendency to the left
This impact which type of error there will be either more FA or more misses
When beta is 1. There is an equal amount of misses and false alarms.
Intersection of distributions of signals and noise is exactly the probability of correct rejection = prob of hits
- B<1, more false alarms than misses.
- Probability of correct rejection > probability of hits
- B>1, more misses than false alarms
- P(hits)>P(correct rejections
Possible to set the optimal level of the RC, given cost, benefits, and signal probabilities
Beta (opt) = P(Noise)/ P(signal) * (V(CR) + C(FA))/(V(hit)+C(miss))
Where beta should be based on the existing ratio of S-N
Slow in asjusting our beta. Beta used by humans is not adjusted to changes in optimal beta.
- Beta (opt) high, human beta is less risky
- Beta (opt) low, human beta less conservative
Sensitivity measure P(A) ~ d'
P(A) = P(hit) +[1-P(FA)] / 2
P(A) = P(hit) + P(CR) /2
increased discrimination (d') makes the curve more curved.
- better training
- amplifying signal vs noise