Learning that depends on the parameters of more than a single stimulus

The notion of'association' is central to both the philosophical and the experimental study of the mind. In philosophy it can be traced back to Aristotle, who proposed that similarity, contrast, and contiguity of images subserve recollection (On memory; Sorabji 1972). 'Associationism', the philosophical doctrine that the mind learns and construes the world bottom-up by associating elementary events, has emerged with British empiricism in the seventeenth century (Warren 1921). Hobbes (1651) talks about 'the train of thoughts' and of 'compounded imagination ... as when from the sight of a man at one time, and of a horse at another, we conceive in our mind a Centaure'. It was, however, Locke (1690) who first used the phrase 'association of ideas', as the title of a chapter in Essay concerning human understanding.

When psychology became an independent empirical discipline towards the end of the nineteenth century, associationism was part of its conceptual heritage. Ebbingahus (1885) was influenced by it when he designed the first quantitative "recall experiments, involving perceptual 'atoms' and their associations. Similarly, Wundt (1896), the founder of the first laboratory of experimental psychology, advocated the study of elementary mental elements and their association in learning, recollection, and thought (Boring 1950a). Over the years the integration of associationism into psychology has also been accompanied by the development of theories2 that kept the centrality of associations yet disposed of the assumption that the mind works solely bottom-up from simple ideas and 'psychic atoms' (e.g. James 1890; Freud 1901; Hebb 1949; Tversky 1977).

Associations play a part in all the faculties of the mind: learning (the formation of new associations, definition 1); recollection (the use of associations as "cues, "priming, "retrieval); and thought (which involves both the generation of new "internal representations, definition 1, and recollection of old ones).3 Here we refer to one aspect only, that of "learning. A popular "taxonomy of learning is based on a dichotomy between 'associative' and 'nonassociative' learning. In contrast with associative learning (definition 2), in nonassociative learning, i.e. "habituation and "sensiti-zation, learning is assumed to depend solely on the parameters of the unconditioned stimulus. Whether in "real-life this is indeed the case, is questionable. Even habituation and sensitization involve associations not only with the history of the subject and its interaction with the stimulus, but also with the "context (Hall and Honey 1989; Rankin 2000). Incidental learning and "insight are occasionally depicted as nonassociative as well, but again, this is a great simplification, as in both cases associations are formed in the mind. Incidental learning involves associations between an input and saliency or motivation. Insight is expected to involve sequential implicit associations of internal representations and their "binding. All in all, therefore, it is possible to conclude that associations of some kind or another are universal, and instrumental in learning in even the simplest organisms and tasks.

The study of associative learning has gained tremendously from the use of animal behaviour "paradigms. At the beginning of the twentieth century two major types of paradigms emerged, which permitted for the first time the investigation of elementary forms of associative learning in laboratory animals, and hence a more "reductive and mechanistic analysis of associations at multiple "levels of analysis. One paradigm was "classical conditioning, associated mainly with Pavlov (1927) and his school. The other was "instrumental conditioning or operant conditioning, associated mainly with Thorndike (1911) and later Skinner (1938) and their schools ("behaviourism). In both types of paradigms, the "subject learns relations among events (definition 1). In classical conditioning these relations are among stimuli, whereas in instrumental conditioning, these relations are among actions and their consequences.

The availability of "controllable protocols of associative learning in animals has provided a fertile ground for the development and test of multiple types of laws and theories of associative learning. These theories differ in the identification of the associated variable and of the principles of association. Main types of associated variables considered in these theories are stimulusstimulus (S-S), stimulus-response (S-R), response-response (R-R), and response-"reinforcer (e.g. see "instrumental conditioning). Stimulus in these theories is commonly an external, sensory stimulus. Note, however, that in definition 2, 'stimulus' is more general and refers to any event that triggers a response in the brain, whether of an external or an internal source, hence it includes also the feedback of motor response. Further, in reality, those are of course not the stimuli themselves that are associated, but rather their on-line "percepts or off-line stored representations. Principles of associations that are considered in theories of associative learning are the frequency of occurrence of the events, their co-occurrence in time and space (contiguity), the probability of linkage (contingency), and the effect or reinforcement (Dickinson 1980; Bower and Hilgard 1981; Mackintosh 1983).

At least in one basic assumption the original British associationism clearly went wrong. This is the depiction of our mental life as dependent only on postnatal associations. Many associations in our brain have innate predispositions. Some authors would even go further to propose that all the associations in our brain are predisposed, and therefore all learning is 'prepared' to some degree or another. This could be due to the existence of certain neural pathways but not others. The generation over time of endogenous pre-representations, which are partially independent of external-world experience but selected by it (Heidmann et al. 1984), could also be constrained by "a priori patterns of connectivity in the brain. An example of a simple type of prepared learning is provided by the form of classical conditioning called a conditioning, in which the modified response is pre-existent. Other examples of prepared associations are "imprinting and "conditioned taste aversion. Whether learning is 'prepared' or not should be taken into account in the search for the cellular and molecular algorithms and mechanisms of learning. For example, presynaptic facilitation of active synapses in the circuit that subserves behaviour (" Aplysia) fits to subserve prepared learning, whereas the activation of silent synapses or the growth of new synapses fit to subserve de novo associations as well.

Selected associations: Classical conditioning, Instrumental conditioning, Priming, Taxonomy

'This definition also fits certain artificial systems, such as smart robots, if 'mental' is construed 'metaphorically. 2As noted in 'algorithm, these are not genuine theories in the mathematical sense of the term, but rather conceptual generalizations. The same is true for 'laws' below.

3For the role of associations in completing memories from partial input in artificial neural networks 'models, see Hopfield (1982), Amit (1989), and Mehrota etal. (1997).

0 0

Post a comment