Signal Detection Theory

MA 4207
 Signal Detection Theory
Roots in World War 2

Radar researchers, Identify Signal from Noise

Used in Experimental Psychology in psychophysical tasks

Windchime Bicycle

Blanket Strings
 Perception and Sensation
• Sensation and Perception
• Stimuli information received by the
receptors: Sensation
• Making accurate inference from
sensory input: Perception

Prior Knowledge->Bias

Perception
Sensation->Sensitivity

• Mismatch=> Illusion
 Signal and Noise

Noisy signal representation, Ds:

Ds = S + N
where S is signal, N is noise

Decode the S from Ds?

 Source of Noise
• External Noise
• Related to noise in the
stimuli
• Interferes with the Signal
• Internal Noise
• Noisy neural response to
Stimuli
• Affects signal representation
• Internal Response
• Decision based on noisy
stimuli
 Noise
• Generally assumed to
be white
• If S is assumed to be
Normal, then Ds also
Normal.
• Typically,
 Sensitivity and Criterion
How well the observer perceives stimuli – or sensitivity

How does the observer choose to respond – or criterion

Methods to assess the two independently

SIGNAL DETECTION THEORY
Biased Unbiased
Criterion Criterion
Sensibility and Criterion
 Sensibility and Criterion
• Sensitivity (Discriminability), d’
• Measure of how close signal and noise
are
• Supposed to be a property only of the
sensory process
• Criterion (Response Bias), C
• Measure of what is considered noise
and what signal
• Susceptible to motivation, strategy,
etc.
• Sensitivity and criterion are
independent
 Example

Internal responses will have some variability because of mixed

neuronal selectivity

Criterion for deciding if there is a tumor present in a scan will

determine classification performance
Criterion shift affects performance
• Conservative criterion → low
false alarm rate, low hit rate
• Aggressive criterion → high hit
rate, high false alarm rate
Computing Sensibility
d’= Separation/ Spread

d’

d’=z(H)-z(FA)
Computing Criterion
Likelihood Ratio

c=-[z(H) + z(FA)]/2
 Sensitivity and Criterion
d' is a measure of sensitivity.
The larger the d' value, the better your performance.
A d' value of zero means that you cannot distinguish trials with the target from
trials without the target.
C is a measure of response bias.
A value greater than 0 indicates a conservative bias
A value less than 0 indicates a liberal bias.
Values close to 0 indicate neutral bias
 ROC
• Receiver operating
characteristic (ROC):
the hit rate as a
function of the false
alarm rate.
• Chance performance
will fall along the
diagonal.
• Good performance
(high sensitivity)
“bows out” towards
the upper left
corner.