Lab Session II:
Signal Detection Theory and Magnitude Estimation
Background:
Purpose and Goals
To illustrate a different way of thinking about
human sensitivity: Signal Detection Theory
To expand your computational horizons.
Signal Detection Theory
Description of Theory
Noise: There are random events in our sensory
neurons that happen randomly. These events
are not caused by external events and are called
noise.
Signal: This is the new name for the stimulus.
Noise and Singal+Noise distributions
So, in looking at a sensory neuron, it
fires even when nothing is present.
Sometimes the neuron fires faster, sometimes
more slowly. Still, no stimulus
(signal) is present. If you plot a curve
where the x-axis is how strong neuron is
firing (often called sensory signal
strength). The y-axis is
probability. If you plot a curve for
how likely any sensory signal strength
occurs when there is not a signal, this is
called the noise curve.
The noise never goes a way, even when a
signal is presented. So, when
presenting a signal to the participant, the
noise will not disappear. There are
still different possible sensory signal
strengths that will occur for a given
signal. Overall, the curve will move
to higher levels on the x-axis (sensory
signal strength axis). Thus, the curve
that is plotted for what can happen when the
signal is presented is called the signal+noise
curve.
Measures:
Sensitivity = d'. The distance
between the two curves (technically in
numbers of standard deviations) is call the
sensitivity. The larger d' is the
easier to tell noise from
signal+noise.
Criterion (one type is called beta).
When the noise and signal+noise curve
overlap, there are some sensory signal
strengths that could be caused by either
noise alone or signal+noise. So the
participant needs to set some criterion
level where below the level, the participant
will say only the noise happened and above
this level, the participant will say that
the signal occurred.
See text Chapter 2 for more information.
Look at it all but concentrate on ROC curves
which I will not review here.
This method argues that there is no such thing as
a threshold. Can you figure out why?
Experimental Method
In signal detection experiments, the stimulus
is only presented on some trials. The
subjects task is to decide if the stimulus has
been presented. This leads to the
following four possible outcomes for each trial
as indicated below:
Adjust criterion until False alarms match
your data
Adjust d' till hits match your data
Read the d' value off of graph
Are your three d's approximately constant for
each relative speed of stimulus (1 pixel or 2.5
pixels)? The three d's come from the three
signal probabilities (.1, .5, .9)
Does the ROC curve do a good job of describing
your data, i.e., does the ROC seem to match your
expectations?
Worth 25 points
Do figures in Excel and place in a word file and
then type the answers to the questions on the same
page. These may not be hand written.
Point: Learn about making graphs and reading them.