Discrimination and categorisation Flashcards
simple discrim
From Pavlov onwards, learning theorists experimented with providing US (in classical conditioning) or reinforcement (in operant conditioning) in the presence of one stimulus (CS+, SD, in general S+), but not in the presence of another (CS-, SΔ, in general S-)
In general, differential responding can be obtained, and we say the animal can discriminate the two stimuli
In early experiments, the stimuli were normally simple, and differed on some obvious physical dimension, e.g. tones of different pitch, lights of different colour
procedures
successive
simultaneous
conditional
successive
present one of the stimuli and see how the animal responds
simultaneous
present two stimuli and see which the animal approaches – normally considered to be easier
conditional
reinforce different responses (or different stimulus-response associations) in the presence of different stimuli
apparatus
Discrimination boxes (mazes with discriminative stimuli added)
Lashley’s jumping stand
Harlow’s Wisconsin General Test Apparatus (WGTA)
Skinner boxes in many variants
Use of colour slides, video and computer displays, and touch screens
key phenom
generalisation
generalisation decrement
generalisation gradient
peak shift
transposition
transfer along a continuum (TAC)
generalisation
some response occurs to stimuli that are physically similar to S+ but not identical to it
generalisation decrement
response to other stimuli is less than that to S+ itself
generalisation grad
a graph relating generalized responding to values on a stimulus dimension
sharpening of generalization gradients when an S- is introduced
- Train S+ on own = broader generalization gradient
peak shift
Responding may be greater to a stimulus other than S+ (S’), on the “other side” of S+ from S- on the stimulus dimension
transposition
If a discrimination between S+ and S- is trained, and then S’ is tested vs. S+, S’ may be chosen
TAC
Training an easy discrimination on a dimension can help the animal acquire a difficult one more than simply practicing that difficult discrimination
peak shift example
Here the effect is demonstrated with naturalistic stimuli, for pigeons with different wavelengths of light, Hanson (1959)
see notes
Spence’s explanation of peak shift
Interacting excitatory and inhibitory generalisation gradients (shown right) produce the result - as long as their shape is chosen correctly.
The theory makes the prediction that peak shift works best with similar (near) S+ and S- (true) and that the shift is greatest in this case (also true).
A modern variant using Rescorla-Wagner has proven very successful
S+ and S- quite strong – take one from the other
see notes
No. of instances of icons B-I present in the stim of exp 1a
see notes
Wills and Mackintosh (1998) an artificial dimension is created by using different icons chosen systematically as shown according to the position on the ‘dimension
S+ and S- overlap
see notes
Their results showed that a good peak-shift could be obtained with an artificial dimension constructed in this way
Humans also show peak-shift, which could have consequences for choice behaviour
see notes
classic theoretical issue: absolute v relative discrim
In any discrimination, an animal learns to respond to one stimulus rather than another. But what is the effective stimulus? Is it absolute or relative?
E.g. does a rat learn to respond to black rather than white or to darker rather than lighter?
For not necessarily good reasons, these two possibilities became embroiled in two complete perspectives on discrimination learning - one derived from early behaviourism, the other from Gestalt ideas and more sympathetic to cognitive interpretations
The supposed crucial experiment: transposition of discrimination to different values on the stimulus dimension
transposition: Wills and Mackintosh (1999)
see notes
Successive – shown one at a time – should respond to darker
Simultaneously – shown at same time but in diff orders
transposition example
An explanation in terms of discrimination on the basis of absolute values
see notes
Hence the mere existence of transposition does not establish relational learning in animals - with the proviso that it must be expected to reverse at extreme values on the dimension.
TAC example
The effect…is that pre-training on an easy problem followed by a shift to a hard problem can be more effective than training on the hard problem only, even when total training times are equated.
This was first reported by Lawrence (1952) using rats.
Choose darker grey
Once shifted to hard problem, instantly better at it
Not due to practice as both had same number of trials
Advantage persists over acquisition
see notes
Training on the easy problem (E+ vs. E-) exploits the bigger difference between the curves for this problem.
Training on the hard problem gives hardly any difference between H+ and H- (lots of generalisation), that’s why it’s hard!
Humans also show a TAC effect, and this could have implications for training phoneme perception
Bigger differences = easier to discriminate
Differences between inhib and excite grads greater
see notes
continuity v non-continuity theory
Is learning a gradual process (continuity) or all-or-none (non-continuity)?
This question is made all the harder when it is realised that a continuity account can be made to look very like a non-continuity account - and vice-versa!
The motivation for the non-continuity account originally came from studying individual pigeons
the Hull-Spence continuity theory (Spence, 1936)
Discrimination of absolute stimuli
A continuity theory: learning occurs gradually
Assumes smooth generalization gradient around the stimuli to which training has actually occurred
Assumes excitatory generalization around S+ and inhibitory generalization around S-. These are hypothetical, internal response tendencies
Observed response tendency is predicted from an (unspecified) monotonic transformation of the algebraic sum of excitatory and inhibitory generalised response tendencies - i.e. excitatory minus inhibitory
With appropriate choices for shapes of the two gradients, this theory can predict transposition, peak-shift and transfer along a continuum.
Krechevsky and Lashley’s non-continuity theories
Discrimination of relative stimuli
Non-continuity theories: learning occurs suddenly
Krechevsky (1932): rats form hypotheses about what is to be discriminated; when they get the right hypothesis, the problem is solved instantly
Predicts position habits, no impact of pre-solution reversal, Transfer along a continuum (Lawrence, 1952)
Fits naturally into modern cognitive ideas about selective attention
compromise theories - combining continuity and non-continuity theory
Discrimination involves both learning what stimulus dimension to attend to, and what stimulus values on that dimension are correct
- Sutherland & Mackintosh (1971) specified that attentional learning is slower to reach asymptote than response learning; allowed attention to multiple stimuli, but assumes that attention is limited so that increased attention to one dimension means less to another.
- This theory predicts the overtraining reversal effect and the impact of overtraining on the relative ease of intradimensional shift and extradimensional shift
assessment of compromise theories
Necessarily weaker (in the Popperian falsificationist sense) than either of the simple theories - More in a theory = harder to falsify But advocates make a good case that multiple factors are involved in discrimination learning
complex discriminations
We’re now moving swiftly up the scale in terms of complexity, but ask yourselves if simple conditioning could explain these abilities.
Following a pioneering experiment of Herrnstein & Loveland (1964), much modern work has concentrated on experiments on discrimination between sets of stimuli
The stimulus sets are usually defined in terms of human concepts, e.g. person vs. non-person, fish vs. non-fish, or artificial concepts defined by specified multiple features
Such categorical discriminations are frequently learned quite quickly
Most discussion has centred on the question of whether animals need to possess concepts in order to perform categorical discriminations - and what it would mean for an animal to “possess a concept”.
2 kinds of abstraction
Perceptual categories – these are all cats: Abstraction = prototype?
- Animals can learn this
Logical categories – these are all fours: Abstraction = concept?
- Animals cannot learn this
diffs of bird visual systems from typical mammalian systems
Cone-rich retinas
Dense receptor matrix over a wide retinal area
Multiple foveas
Classes of cone differ by oil-droplets filtering light, not by visual pigment
More than 3 types of cone
Spectral brightness response and discrimination
High flicker fusion frequency
Ectostriatum rather than visual cortex
some special features of the pigeon visual system
Two foveas in each eye, one forward (binocular), one lateral
Two visual systems have different functions and psychophysical responses
Very wide range of view
U/V light detected, and affects colour matches
Plane of polarisation of light discriminated
perceptual categories
Herrnstein and Loveland (1964) “Higher order concept formation in the pigeon”
Pigeons learned to peck in the presence of a picture of a person, and withhold pecks in the presence of a picture with no person in it
Stimuli (holiday slides) varied greatly in number of people, posture, whole/part person, clothing, etc
After successful learning, transfer trials show correct response to new stimuli
Could just learn the whole picture rather than just looking for the people
But then when slides mixed still performed above chance – learn the concept of person
continuing work on perceptual categories
Other concepts, e.g. fish, leaves, trees, cats, dogs, male/female human faces, Bach vs. Stravinsky, etc
Some concepts that are ecologically valid for the species, e.g. individual conspecifics, prey items, locations.
theories of category discrimination
Rote learning or absolute discrimination:
- Claimed to be ruled out by successful transfer to new instances
- but what if there was also stimulus generalisation?
Multiple linear feature model:
- Predicts a superreleaser (prototype) effect that does not always occur (but on the other hand often does…)
- Often difficult to demonstrate control by multiple features
Configural (Exemplar) models
testing the theories: artificial polymorphous categories
see notes
Dennis et al 1973: What makes a stimulus a member of group A?
The difference is that group A are more symmetrical and group B aren’t
People find this more difficult than pigeons do
Fersen and Lea (1990): multiple complex features
see slides
All features controlled behavior (eventually, after special training)
Reversal on a subset of stimuli transferred to other stimuli (instance to category generalization)…
…but NOT to other features…
…so no evidence of a coherent concept
Multiple linear feature model describes data well - basically a Rescorla-Wagner model using units to represent features will do the trick.
Pigeons don’t change opinion about other features even if learnt the change
Humans can learn this
same v diff discriminations
Matching to sample tasks and oddity from sample tasks (Zentall & Hogan 1974, Wright et al 1988, and many others): responding based on identity/difference if enough exemplars are used
Evidence: Transfer to novel stimuli on the first trial.
- Has been found in Chimps, dolphins and corvids. Was eventually obtained (though only after some considerable effort!) with pigeons; Colombo, Cottle and Frost (2003).
- May have learnt something about the task
Implication: The animals have the concept of same vs. different?
- Consistent with this but often also susceptible to the interpretation that they are discriminating on the basis of recency – sense of familiarity rather than a concept of same v different
Sees sample image
Small delay (sometimes 0)
Get comparison
Pick matching stimuli
Novel stimuli – e.g. blue and yellow circle
