I have a piece in the Guardian, ‘The science of learning: five classic studies‘. Here’s the intro:
A few classic studies help to define the way we think about the science of learning. A classic study isn’t classic just because it uncovered a new fact, but because it neatly demonstrates a profound truth about how we learn – often at the same time showing up our unjustified assumptions about how our minds work.
My picks for five classics of learning were:
- Bartlett’s “War of the Ghosts”
- Skinner’s operant conditioning
- work on dissociable memory systems by Larry Squire and colleagues
- de Groot’s studies of expertise in chess grandmasters, and ….
- Anders Ericcson’s work on deliberate practice (of ‘ten thousands hours’ fame)
Obviously, that’s just my choice (and you can read my reasons in the article). Did I choose right? Or is there a classic study of learning I missed? Answers in the comments.
Mind Hacks 
Is Pavlov simply too meta-classic? 😉
John Patrick,”Training: Research and Practice” is a particularly good summary of the development and use of training to improve performance.
Shanteau, “Competence in Experts,” is an outstanding discussion comparing and reconciling the biases lit and the expert performance lit. Also the related studies reviewed in his edited collection …
Click to access 2002035181.pdf
Gary Klein, “Sources of Power” and other work on naturalistic expertise and the macrocognition framework.
This is quite a terrible collection that teaches very little about how learning works, even ignoring that it’s mostly about memory (without making the critical statement that memory really is learning).
In particular it suggests that learning is associative, when we have known for fifty years now that learning is a discriminative process that tunes an organisms expectations to the experiences it makes and thereby reduces its uncertainty about the world – by eliminating uninformative connections rather than merely building up new ones (there’s a reason why you’ll never have as many synapses as you do when you’re a baby).
Instead, for example, you could have started from the foundations of Pavlov and Thorndike, continued to Skinner and then focused on Rescorla&Wagner and Kamin and how virtually every learning model today is discriminative and not associative, since that is the most common misconception among laypersons as well as psychologists. Or to say it in someone else’s words:
“How often have I said to you that when you have eliminated the impossible, whatever remains, however improbable, must be the truth?” – Sir Arthur Conan Doyle, Sherlock Holmes: The Sign of Four (1890)
You want to add pavlov, rescorla-Wagner, and Kamin, so people don’t get the impression that learning is associative? Odd…
It’s only odd if you don’t understand the difference between association and discrimination that was the whole point of my comment. Pavlov and Thorndike are of course only there for historical context and to pave the way for Kamin and Rescorla.
Rescorla, Robert A. “Pavlovian conditioning: It’s not what you think it is.” American Psychologist 43.3 (1988): 151.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.294.1082&rep=rep1&type=pdf
Pavlovian and operant conditioning are both a matter of the effects of contingencies on the future probability of behavior. Period. The anachronistic, mentalistic language of associationism is irrelevant. And “discrimination” is, itself, a matter of exposure to discriminative contingencies, though it can no longer be attributed to just a summation of extinction and reinforcement. It can be pretty cogently argued that discrimination involves the establishment of what is now called “observing behavior.” And nothing is gained by interposing the mentalistic language of “expectation.” Oh, BTW, Rescorla is an associationist mentalist. Also, BTW, nonhumans will not respond if the consequence of responding is production of an S-delta, but will respond if it produces an Sd, even though both consequences each provide one bit of information (i.e., “uncertainty reduction”). Finally, the physiological mediation of “learning” is a different field from the analysis of behavior qua behavior which is an independent science as shown by the enormous (but often overlooked) success of behavior analysis.
I don’t have a strong opinion on the debate of whether there are representations that allow for expectations and predictions or not, although there is evidence that behaviour driven by an internal model looks different from model-free learning, e.g.
Daw, Nathaniel D., Yael Niv, and Peter Dayan. “Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control.” Nature neuroscience 8.12 (2005): 1704-1711.
Gläscher, Jan, et al. “States versus rewards: dissociable neural prediction error signals underlying model-based and model-free reinforcement learning.” Neuron 66.4 (2010): 585-595.
and non-representational accounts of, say, embodiment (Wilson, Golonka and colleagues), however elegant, never seem to have an actual learning story that could be implemented. But it doesn’t quite matter with respect to whether increasing association strengths between things that occur together is enough to produce learning, which, and from you answer I take it that you agree here, it isn’t.
I don’t care what Bob “is”, the framework is discriminative. Calling the language used irrelevant doesn’t make any sense, given that I disagree with the message that is being communicated with it. Similarly, divorcing the implementational level (“physiological mediation of learning”) from the behavioural one seems downright ignorant.
I don’t understand your point on different human responses given the same amount of information, please elaborate on that with a reference. Also on “observing behaviour” and how it explains discrimination learning, which I don’t understand either.
@anonymous
I accept that describing Skinner’s result as demonstrating the power of associations is somewhere in between unfortunate and plain wrong.
So, a question for you, would the relevant section be acceptable to you if I replaced all instances of “associations” with “discriminations” in the original piece?
Tom: I accept that describing Skinner’s result as demonstrating the power of associations is somewhere in between unfortunate and plain wrong.
GS: This is a misrepresentation of Skinner’s position. Skinner’s view is that contingencies are responsible for Pavlovian and operant conditioning (different sorts of contingencies of course). And that is simply a fact. Associationism is, essentially, the view that conditioning is the result of associations between internal representations and Skinner, of course, never held that view. And even if “representations” were not discussed the “association” was “internal.” But, as Skinner always said, the animal doesn’t “associate” anything – the stimuli (Pavlovian)or the response–>consequence contingency (operant) exists in the environment. It is the researcher who associates these things in the sense of arranging contingent relations. BTW, I’m guessing you have never read much by Skinner. Right? That is typical, anyway. No one likes to bad-mouth behavior analysis and behaviorism more than those that never actually read anything – Chomsky being a perfect example!
A: I don’t have a strong opinion on the debate of whether there are representations that allow for expectations and predictions or not, although there is evidence that behaviour driven by an internal model looks different from model-free learning, e.g.[…]
Daw, Nathaniel D., Yael Niv, and Peter Dayan. “Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control.” Nature neuroscience 8.12 (2005): 1704-1711.
GS: I guarantee you that “driven by an internal model” is an assumption like all of the other explanatory fictions reified into existence by cognitive “science.” Oh…I see, the paper you cited is a modelling paper. Which is OK…but that means the existence of an “internal model” is an assumption.
A: Gläscher, Jan, et al. “States versus rewards: dissociable neural prediction error signals underlying model-based and model-free reinforcement learning.” Neuron 66.4 (2010): 585-595.
GS: This is typical stuff for cognitive neuro“science.” When, for example one visual pathway is differentially active depending on what a stimulus is and another is differentially active depending on where in the visual field a stimulus is, it is said that that these are separate parts of a “representation” that (giggle) must be recombined before they are seen (presumably by the vision homunculus – not to be confused with the “executive” homunculus who everyone knows is named Homey and he lives in the PFC). And no doubt the “representations” must first be inverted since the retinal image is up-side down (LOL). An alternative notion is that “telling what a thing is” and “telling where a thing is” are different responses and the aforementioned activity is part of those responses. Behavior, including perceptual behavior, is the functioning of a large system – that system has parts which, by themselves are probably best not called “behavior” but are part of the overall functioning identified as “behavior.”
A: […]and non-representational accounts of, say, embodiment (Wilson, Golonka and colleagues), however elegant, never seem to have an actual learning story that could be implemented.
GS: Ahh…since you cannot build an artifact that mimics behavior (perceptual or otherwise), conceptual views that deviate from yours must be wrong. But, of course, “representationalism,” “information-processing” etc. are the assumptions that underlie attempts to simulate behavior – indirect realism comprises by far the most accepted underlying assumptions. “Emobodiment” (poorly described but still closet behaviorism) does not, by definition, fit with these assumptions and, hence, the modeler who has only the tools of representationalism is stymied.
A: But it doesn’t quite matter with respect to whether increasing association strengths between things that occur together is enough to produce learning, which, and from you answer I take it that you agree here, it isn’t.
GS: Red herring. Conditioning, operant or classical, is a function of contingencies (though the temporal relation among elements remain important). How this is “implemented” is up for grabs. But the assumption that the brute fact of contingencies over temporal contiguities means that there is a “representation” is just that – assumption. And, unfortunately, assumptions are neither purported (verifiable) facts, nor are they hypotheses.
A: I don’t care what Bob “is”, the framework is discriminative.
GS: The framework is that of contingencies. Discriminative behavior is a function of discriminative contingencies.
A: Calling the language used irrelevant doesn’t make any sense, given that I disagree with the message that is being communicated with it.
GS: Not sure what you are driving at here. Arranging differential contingencies gives rise to differential behavior in the presence of the different stimuli. Contingencies comprise the primary independent-variable and discrimination is an outcome.
A: Similarly, divorcing the implementational level (“physiological mediation of learning”) from the behavioural one seems downright ignorant.
GS: Yet it is one that has been enormously successful in building a science of behavior. And if behavior analysis hadn’t been so vigorously misrepresented by the early cognitivists, more mainstream psychologists and those they have corrupted might know about that. And there is precedent for keeping a level separate from the reductionistic level, at least for a time. This allowed the development of chemistry and thermodynamics. Skinner’s position has always been that a science of behavior qua behavior must proceed before a neuroscience explanation can be obtained for behavioral facts are what are to be explained at the reductionistic level.
A: I don’t understand your point on different human responses given the same amount of information, please elaborate on that with a reference. Also on “observing behaviour” and how it explains discrimination learning, which I don’t understand either.
GS: I wasn’t referencing humans – they sometimes do not show the effect I was describing. But, then, all the humans looked at had extensive verbal repertoires. Briefly, in a simple observing experiment, say with pigeons, key-pecks are occasionally reinforced in the presence of one stimulus (say, a green key-light) and not at all in the presence of the other (say, red). The result is, of course, that the pigeon tends to peck at high rates during green and little or not at all during red. The differential stimuli are then replaced with a single stimulus but the reinforcement and extinction schedules continue to alternate unpredictably. However, the differential stimuli may be briefly reinstated if the pigeon completes some requirement on a different key (the observing key). That is, completion of the observing requirement results in the brief presentation of the stimulus previously correlated with the schedule in effect. Under these circumstances both S-delta (extinction) and Sd (reinforcement) provide 1-bit of information. However, if S-delta is the only consequence of observing responding (if the food schedule is currently arranged, observing responses have no effect), then pigeons will cease responding. However, if the Sd is the only consequence of observing then observing responding will be maintained. Same amount of information, different outcomes…it is not information per se that maintains responding, at least in non-humans (and this reflects the basic processes uncontaminated by humans extensive verbal history). I have included links to two papers, one empirical and one conceptual. The former is a bit more complicated than the simple one I described, but it is of the same ilk:
http://onlinelibrary.wiley.com/doi/10.1901/jeab.1974.22-525/abstract
http://onlinelibrary.wiley.com/doi/10.1901/jeab.1985.43-365/abstract