Modelling Cognitive Processes Flashcards
Why do we need modelling?
Cognition = noisy ie
Searching for signal in noise - targets from distractors
Memory/perception/etc tasks - how to detect the target? how easy is it? does it stand out?
Decision tasks - when to say yes vs no
What does modelling help us understand?
Separates decisional from non-decisional processes - what affects the former or latter?
Helps interpret scientific data, distinguish between theories
Helps compare participants and controls
Helps us learn/improve/optimise the science and our own cognitive performance
What are three prerequisites for modelling?
Need to have a good understanding of how the cognitive process operates
Need to understand what a particular modelling method requires
Need to have meaningful data to input and an appropriate way of processing it (in order to get meaningful output)
What is psychophysics?
The relationship between stimuli in the environment and sensations
Assumptions similar those underlying nearly all experimental psychology - behaviours help us infer mental states
Close links with neuroscience ie neural mechanisms of vision and their links to behaviour
Gustav Fechner
What is Detection Theory? (Signal detection theory)
WW2 radar origins - enemy plane or something innocuous? Important to know to minimise death/waste of resources - correct hits (H) and correct rejections (CR) vs false alarms (FA) and misses (M) - maximise first two, minimise second two
Optimise: discrimination = what is signal vs noise?
Criterion setting = amount of evidence needed to say yes
How do you calculate discrimination?
Hit rate (HR) - probability of saying yes when a signal is presented = H/(H+M)
False alarm rate (FAR) = probability of saying yes when noise is presented = FA/(FA+CR)
What is a z score?
How many standard deviations +/- from the mean of a normal distribution (mean at 0) is the HR and FAR away
What is d’?
A measure of discrimination
d’ = z(HR) – z(FAR)
The higher the d’ = greater the distance between distributions and the greater the discriminatory ability
What is c?
A measure of bias; the distance between the criterion and the intersection of the distribution
c = -0.5 [z(HR) + z(FAR)]
The higher the c, the less a person is willing to say yes
What is the relationship between d’ and c? (discrimination and bias)
Thought to be independent of one another; though some debate as to whether this always holds
What are three other assumptions necessary for detection theory?
That d’ and c are independent
That the distributions (of H/M/FA/CR) are normal (though other models exist for nonnormal distributions)
The evidence that feeds into the dimension is graded
What can vary in a detection theory distribution?
Signal and noise distributions - equal variance signal detection models assume both distributions are the same; unequal variance models assume one has greater variance than the other
Unequal models - eg in recognition memory - variance of target/signal = greater than the lure/noise because targets can vary so much; requires d(a) instead of d’
What relevance does detection theory have to real life?
Use in eye witness testimony to establish best protocol for presenting suspects ie sequential vs simultaneous procedures
What should all lineups be? (and not be?)
Fair - all people presented are reasonably similar to one another
Accompanied by unbiased instruction - no forcing of choice, informing that culprit may not be present
Should not be a showup ie presenting with one face and asking if they are the culprit
What possible decisions can be made when choosing from a line up?
- Picking a person:
correct = culprit
incorrect = person is an innocent suspect
incorrect = person is a foil/not a suspect - Rejecting the line up:
correct = when person is an innocent suspect
incorrect = when person is the culprit
Detection theory modelling can be used to optimise the correct rejections/identifications and minimise errors
What does modelling suggest the best lineup type is?
Historically - simultaneous = worse (higher H but also FAs) - people less able to distinguish between seen/unseen lineup members = worse discriminabiltiy
Used to use a ‘diagnosticity ratio’ but this confounded criterion with discriminabiltiy (how conservative/liberal someone is with choosing vs how good someones memory is)
Different methods (Mickes/Flowe/Wixted 2012) found better discriminabiltiy with simultaneous line-ups under new methods - though contested
Basically - the choice varies with policy, culprit walking free vs innocent person being convicted - liberal vs conservative criterion
What relevance is detection theory to optimising performance?
Negative scoring on exams:
Assumption - penalties for guesses as they reflect luck rather than knowledge
However - guesses can reflect partial knowledge and metacognition (abilities to evaluate own knowledge and use of strategy) meaning final scoring is no longer an assessment of knowledge
What are dual vs single process models of recognition memory?
Dual - familiarity and recognition are distinct processes; familiarity is graded so can be modelled using a SDT-related tool; recollection is all or none
Single - all memory processes feed into one memory evidence dimension
What are Receiver Operating Characteristic (ROC) curves?
SDT-related tool for theory testing
Binary responses (yes/no, old/new etc) can be scaled using confidence ratings ie moderately confident yes or very confident yes etc - for each point on the scale, you can calculate HR and FAR for all answers with that level of confidence or higher - HR and FAR pairs are at different levels of bias and can be plotted on a curve
(I dont know what the fuck i’ve just written here, should have gone to this lecture…)
How do you interpret ROC curves?
Area under the curve (AUC) = measure of discrimination
Larger the area = better people discriminate between signal and noise
Points on the curve = criteria for subsequent responses ie criterion for very confident no or higher
Show how discrimination varies at different levels of bias (criterion setting) - how people are willing to say yes
What are zROC curves?
HRs and FARs can be converted into zROCs which provide info about the ratio of variances of the two distributions (equal variance = slope of 1; unequal = different to 1) and whether the distributions are normal (normal zROCs = linear; non-normal = curvilinear
What are some examples of zROC curves?
Dual process models of recollection and familiarity:
Yonelinas et al (2002) - hippocampal lesion studies:
Hippocampus implicated in recollection - controls = curvilinear zROCs stemming from a recollective process shown by the line bending with the most confident responses compared to the lesion patients who showed a linear response
However - memory strength can affect zROC shape: weaker = linear, with a slope of 1 (equal variance); stronger = increased and unequal variance ie many strong memories and people make unnecessary distinctions between moderate and very certain yes which produces an artefact: curvilinear zROCs
What is a diffusion model?
Ratcliff, McKoon et al; came after SDT but also deals with signal detection
Computationally complicated but captures data complexity better ie better at dealing with temporality
What are some assumptions and limitations of diffusion models?
Assume evidence accumulation over time = noisy process
Separate the quality of evidence from the criteria and from non-decision processes ie stimulus encoding or response execution
Limited as are mostly applicable to fast decisional processes
How do diffusion models work?
z is the starting point of data accumulation; a and 0 are criterion which data accumulates towards, when reached a response is initiated; v is the drift rate/rate of accumulation and represents the quality of evidence
a, 0, z and v vary across tasks/conditions - criteria can move towards/away from starting point concurrently or separately
Starting point can move up or down depending on initial bias
Drift rate varies and variability tends to be normally distributed
Drift rate criterion = what is considered old and what new?
What can be modelled using diffusion models?
Effects of ageing on memory and lexical decisions
Single neuron recordings in brightness recognition tasks in macaques
ERP data etc
ROCs can also be produced from diffusion models by manipulating decision boundaries +/- drift rate criterion
