|
Psychology of Communication: learning
|
|||
|
Psychology of Communication: learning
|
Until the 1960s behaviourism was the dominant, virtually the paradigmatic, approach to learning. Around that period it gradually became increasingly apparent that there were aspects of learning that it was unable to account for.
The anomalies evident from behaviourists' own observations of 'latent learning' were proving increasingly difficult to account for. They had observed that rats left to their own devices in a maze would find their own way around even when no reward such as food was present, nor even a stimulus such as the scent of food. Behaviourists attempted to save the appearances by postulating a 'curiosity drive' which was 'reduced' by exploration. As Dennett comments,
as an attempt to save orthodox behaviourism, this was manifestly vacuous, but that does not make it a hopeless idea in other contexts; it acknowledges the fact that curiosity - epistemic hunger - must drive any powerful learning system
Dennett (1996 : 92)
Skinner himself in Verbal Behaviour vacillated between behaviouristic explanations and mentalist explanations which were forced upon him, for example in people's use of similes or metaphors. In particular, his daughter's observation that soda pop tasted 'like my foot's asleep', whilst providing Skinner with an observable reaction to a stimulus, is at the same time compared with an experience to which he has no access. He accounts for this by referring to the sensation of pins and needles as a 'private stimulus'. His claim that such stimuli play a large part in the generating 'verbal behaviour' leaves him wide open to Chomsky's accusation of closet mentalism. (Walker (1984 : 123 et seq.)) The thirty-year old Chomsky's devastating review of the highly-respected Skinner's Verbal Behaviour concluded that Skinner was 'play-acting at science'. Though a scientific revolution never happens at a particular time and place, Chomsky's attack was certainly a major step in the 'cognitive revolution'.
A major source of embarrassment for behaviourism was also its inability to define 'response'. Observations revealed that the actual topography of an organism's response varied significantly from one trial to the next. There appeared to be no specific response due to the acquisition of an association. All that the responses had in common was what they achieved - an observation which cries out for a mentalist explanation. (see Amiri (undated) : 23).
Hypothesis-testing, which had previously been rejected as overly mentalistic, was presented around this time as part of a rigorous mathematical model and thus became more acceptable, introducing an intentionalist approach into learning theory in addition, or opposition, to associationism. In Dennett's terms, it became evident that, in addition to being Darwinian and Skinnerian creatures, we are also Popperian creatures, having the ability, in Popper's felicitous phrase, to 'permit our hypotheses to die in our stead'. Clearly, if we accept that hypothesis-testing takes place, then the intentionality which that implies entails the existence of some kind of inner environment which contains information about the outer environment - which clearly takes us well beyond behaviourism's insistence on mere observation of behaviour to the consideration of mental representations of the world.
Around the time of Chomsky's article early researchers in artificial intelligence were demonstrating that computers could decipher codes and produce proofs. If it was possible to approach machines as if they were 'internally' manipulating symbols, then why not approach human beings in the same way?
The cognitivist approach to the mind has in common with the approach from Chomskyan linguistics that it is concerned with symbolic representations and rule manipulation and modification. On the other hand, it has in common with behaviourism that it is resolutely empiricist, unlike the 'rationalist' approach influenced by linguistics. Much of early cognitivist development shows the influence of the work of the Gestalt theorists who, unlike behaviourists, placed great emphasis on the importance of organizational processes in perception and learning. (for some brief comment on gestaltism, see the section on gestalt approaches to perception). Jean Piaget's studies in learning development also had considerable influence on cognitivism, notably on the notion, also drawn from Gestalt theory, that knowledge is organized and structured, as well as the view that, for learning to occur, it must be incorporated within existing cognitive structures, from which it follows that new experience and prior knowledge must overlap. Of importance also, especially to theorists of human learning, is Piaget's emphasis on four distinct stages of cognitive development, each characterized by qualitatively different forms of thought.
A significant influence also came from Chomskian linguistics, notably his rejection of finite state grammars in favour of transformational generative grammars and the concomitant rejection of behaviourist approaches to learning which
have not provided any way to account for or even to express the fundamental fact about the normal use of language, namely the speaker's ability to produce and understand instantly new sentences that are not similar to those previously heard in any physically defined sense or in terms of any notion of frames or classes of elements, nor associated with those previously heard by conditioning, not obtainable by them form any sort of 'generalization' known to psychology or philosophy.
Chomsky (1965 : 57-58)
Chomsky roots his linguistics firmly within the rationalist tradition of Descartes, the Port-Royal Logic, Leibniz and others, drawing attention specifically to Humboldt's view that one cannot really teach language, but only 'present the conditions under which it will develop in the mind in its own way' (Chomsky (1965 : 51)). (see Humboldt (1963/1827-1829 : 191 et seq), in which he argues the Platonic view that learning is re-learning or reminiscence; also Chomsky's comparison of Plato's notion of reminiscence from an earlier existence with modern biology's notion of genetic endowment, formed over millions of years of evolution (Chomsky (1996 : 10)), which would include, amongst other faculties, the Chomskian language acquisition device. Note that Chomsky and his close colleague Jerry Fodor insisted from the very beginning on a modular view of the brain's structure. Chomsky's main line of attack was certainly against behaviourism, but he also attacked the non-behaviourist Piaget, who insisted on viewing language acquisition as a function of general intelligence. Chomsky and Fodor would have none of this and their view of the mind as modular, consisting of a number of specialized devices or mental 'organs', is now widely accepted (though there are certainly challenges to this orthodoxy, including intriguing suggestions of a possible hybrid view, incorporating both Piagetian developmentalism and Chomskian modularity). As a layperson, I would recommend Stephen Pinker's The Language Instinct (1994), perhaps the most convincing and persuasive work deriving from Chomskian linguistics, demonstrating its practical application to the understanding of language, as well as its potential for forming the basis of a broader understanding of how the mind works. For the student of cultural studies, I would suggest, many of Pinker's politically incorrect and unfashionable suggestions may prove a salutary challenge.
'[…] cognitive psychologists have consistently found that knowledge is organized. There is some type of structure in knowledge' (Amiri (undated)). Thus, the cognitivist approach emphasizes the cognitive structure which the learner forms in memory. This structure preserves and organizes the events which occur during the learning situation. In a test situation the learner scans the test stimulus against her memory to determine appropriate action. The action performed depends upon the cognitive structure retrieved from memory and the context of the test. The learner's response is therefore the result of a decision process, which will depend upon the subject's memory for prior events, as well as upon the test situation.
Our ability to generalize is the best evidence that we use many mental representations. These, together with rules on their combination and transformation are capable of generating an inexhaustible supply of further representations. This notion particularly underlies the computational theory of the mind, which is dominant in cognitive science. According to Pinker (1998/1997 : 91), natural selection has shaped our 'modular, multiformat' minds on the following computational principles:
The computational view of the mind is summarized by Stephen Harnad thus:
Symbolists emphasize that the symbolic level (for them, the mental level) is a natural functional level of its own, with ruleful regularities that are independent of their specific physical realizations. For symbolists, this implementation-independence is the critical difference between cognitive phenomena and ordinary physical phenomena and their respective explanations. This concept of an autonomous symbolic level also conforms to general foundational principles in the theory of computation and applies to all the work being done in symbolic AI, the branch of science that has so far been the most successful in generating (hence explaining) intelligent behaviour.
Harnad (1990)
The computational model of the mind is famously contested by Searle, whose 'Chinese room' thought experiment is used by him to demonstrate that syntax does not produce semantics and symbol-manipulation in itself does not constitute thinking. Searle's argument is often construed merely as a dispute about the possibility of AI. Indeed, his own presentation of the argument often suggests a biologist bias and often smacks of Cartesian dualism (it is not as evident to me as it appears to be to Searle that a computer made our of 'beer cans strung together with wires and powered by windmills' should necessarily be incapable of an activity I might recognize as thinking). Nevertheless, although Pinker (1998/1997) dismisses the Chinese Room, Chinese Gym, Chinese Robot and so on as mere arguments about the meaning of the word 'understand', many cognitive scientists consider Searle's argument raises serious questions about the fundamental basis of their science, with its emphasis on 'syntax' . In particular it raises the so-called symbol-grounding problem, which has exercised Stephen Harnad for some time. (see Harnad (1990))
An increasingly important challenge to the symbolic approach comes from connectionism, which may be classified by some as a cognitivist approach (e.g. Huitt (1998)), by others as associationist and by yet others as a hybrid (e.g. Bechtel and Abrahamsen (1991).
By the 1980s the limitations of rule-governed, knowledge-based expert systems for the further development of AI were becoming apparent, despite obvious successes such as MYCIN (which proved successful in producing medical diagnoses) and Prospector (which was useful in locating mineral deposits). In the study of human cognition, rule systems were becoming increasingly ad hoc, complex, inflexible and brittle. Coupled with the insights becoming available from neuroscience, connectionism appeared to some to offer a way around such difficulties in the study of human cognition and held out the possibility of domain-independent and context-sensitive machine learning in AI.
Some, notably Chomskian linguists, see the symbolic rule-governed approach as a defining characteristic of cognitivism and thus see connectionism as something other because it rejects the necessity of rules (see Dror and Dascal (undated) and Bechtel W and Abrahamsen A (1991): 255 et seq. for a discussion of the relationship between connectionism and the other disciplines of cognitive science). Pinker (1998/1997: 112 et seq), for example, argues that a neural network alone is insufficient to account for language learning. In his view, the structuring of networks into programs for manipulating symbols is vital to human intelligence and he argues strongly from the evidence of already existing neural network applications that they are fundamentally inadequate to the tasks set them.
Connectionism's statistics-based adjustment of 'weights' and excitation or inhibition of neurons is seductively analogous to the insights of neuroscience into, for instance, what is known as long-term potentiation. (Long-term potentiation: cell 1 receives a stimulus, which causes it to fire. If it fires it fast enough, its neighbour, cell 2, will also fire. Cell 2 is changed by the process since receptors normally hidden within the cell are brought to the surface. As a result, the cell remains more responsive to its neighbour for some hours, or even days. If cell 1 fires again during this period of receptivity, only a weak firing is necessary to activate cell 2. Every time the two cells fire together, the link is strengthened until they become permanently bonded. When a group of linked neurons fires, it triggers a memory - Carter (1998: 176)) Connectionism's distributed representation of concepts also appears to be a parallel to the way that the brain actually processes much information.
But Pinker insists that the structuring process of specialized mental modules is essential if the network is to approximate to the performance of human intelligence (though he possibly rather overstates his case since modularity is increasingly being incorporated into connectionist models). For example, Hinton devised a three-layer network to compute family relationships. He drew attention to its ability to generalize to new pairs of kin; but it transpired that it had to be trained on 100 of the possible 104 pairs in order to generalize to the remaining four, and each pair had to be fed into the network 100 times:
The numbers are typical of connectionist networks, because they do not cut to the solution by means of rules but need to have most of the examples pounded into them and merely interpolate between the examples. Every substantially different kind of exmaple must be in the training set, or the network will interpolate spuriously, as in the story of the statisticians on a duck hunt: One shoots a yard too high, the second shoots a yard too low, and the third shouts 'We got him!'
Pinker (1998/1997: 130-131)
Behaviour modification
Berlo on learning
constructivism
Conditioned Reflex
Schramm's Fraction of Selection
Social learning theory
|