The occurrence of a learned response in circumstances that differ from those prevailing during acquisition

The functional counterpart of 'generalization' is 'discrimination', which is the ability to distinguish one stimulus or response from another. This is why it is convenient to treat these faculties together. Discrimination increases the repertoire of fine-tuned "perception and response, but places a burden on "capacity; generalization diminishes the sensitivity to input noise, but limits the degrees of freedom of the behavioural repertoire and could increase the proportion of false-positive responses. A "subject could be well capable of discriminating among two events, but still generalize its response to them if this pays off behaviourally. The optimal balance between discrimination and generalization is a function of the task and the situation. This balance in nature is nicely illustrated in the primitive "fear conditioning reflex (LeDoux 1996). Suppose we unexpectedly detect a snake-like shadow. The visual information travels to the "amygdala, a brain centre involved in the emotional response, either directly via the thalamus or indirectly via the "cerebral cortex. The thalamo-amygdalar pathway transmits only some general features of the stimulus (that it is a long undulating object), but not detailed sensory attributes (revealing that it is only a loose black water pipe). This channel makes it possible for the organism to respond faster to the generalized gestalt of a snake, risking a false-positive but at the same time minimizing the probability of a fatal false-negative response. The slightly slower thalamocortico-amygdaloid route provides information about the detailed, discriminative sensory attributes of the stimulus ("cue), and prevents further costly physiological and behavioural reactions.

Generalization, discrimination, and their modulation by experience became fashionable research topics early in the history of psychology (Boring 1950a; Keller and Schoenfeld 1950). Since then, it is common practice to distinguish 'stimulus generalization' from 'response generalization'. Stimulus generalization refers to situations in which a subject that had learned to respond to a particular stimulus comes to elicit the same response to other, similar stimuli. Response generalization refers to situations in which a subject that had learned to respond to a given stimulus comes to elicit other, similar responses to the same stimulus. A useful measure of stimulus generalization is the 'generalization gradient', which depicts the range of response to a conditioned stimulus (Figure 32). In some paradigms, the more difficult the task, the less the generalization, hence the steeper the generalization gradient (Ahissar and Hochstein 1997). A corresponding measure in discrimination is the 'just-noticeable-difference' ('jnd'). This is the smallest change along a stimulus "dimension that can still be discriminated. The relationship between the change in stimulus intensity that can just be discriminated (A^) and the intensity of the stimulus is approximated by 'Weber's law': A^/^ = constant. The relation between the stimulus magnitude and the subjective sensation magnitude (y) is given by 'Fechner's law': y = klog ^ (where k is a constant; Gescheider 1997). Note that both are not really 'laws' but merely generalizations, which only approximate reality under limited conditions ("algorithm).

Discrimination is invaluable in the study of conditioning paradigms, in which the subject's ability to acquire and store new information is quantified by its ability to distinguish among stimuli and associate the appropriate stimulus with the unconditioned stimulus or "reinforcer ("classical conditioning, "instrumental conditioning). The use of discrimination protocols in probing the functional neuroanatomy of visual cortex and its role in learning in the "monkey is but one example (e.g. Mishkin 1982). A similar use in dissecting the brain systems that process and learn chemical information in the "rat is yet another example (e.g. Schul et al. 1996). An important variable in such experiments is the sensory attributes of the objects to be discriminated. Those attributes that support sensory differentiation are traditionally termed 'discriminanda' (Tolman 1932). Appropriate discriminanda, e.g. of food items as

Pigeons Key Generalization

Fig. 32 Generalization gradients in the pigeon. Four groups of pigeons were trained to peck at an illuminated key to get food *reinforcement. One group was trained to respond to illumination at 530 nm, another at 550 nm, still another at 580nm, and the fourth at 600nm. Generalization testing was carried out by presenting the pigeons with random series of lights at different wavelengths and measuring their key pecking as a function of the wavelength. The curves depict the range of response for each group. The same graphs can, of course, be treated as discrimination curves. (Adapted from Guttman and Kalish 1956.)

Wavelength (nm)

Fig. 32 Generalization gradients in the pigeon. Four groups of pigeons were trained to peck at an illuminated key to get food *reinforcement. One group was trained to respond to illumination at 530 nm, another at 550 nm, still another at 580nm, and the fourth at 600nm. Generalization testing was carried out by presenting the pigeons with random series of lights at different wavelengths and measuring their key pecking as a function of the wavelength. The curves depict the range of response for each group. The same graphs can, of course, be treated as discrimination curves. (Adapted from Guttman and Kalish 1956.)

opposed to nonsense objects, can markedly facilitate discrimination learning (Jarvik 1953). Discrimination analysis is routinely used in psychophysics to study perceptual competence and use-dependent modification in perceptual "skill (Green and Swets 1988; Gescheider 1997). Generalization is also used to infer the ability of animals to extract rules and in assessing the phylogeny of cognitive abilities (Harlow 1949; Wright et al. 1988; Slotnick 1994; "learning set; for "classics on human ability to generalize and categorize, see Rosch et al. 1976; Tversky 1977). Indeed, simple generalization is considered as a rudimentary form of categorization, induction and concept formation. Deficits in generalization could result in marked behavioural disadvantage, and in humans impair "performance on even mildly demanding jobs ("capacity, "mnemonics). Overgeneralization is also no good, as it could lead to inaccuracy of response, and, under certain conditions, even worse, to generalized anxiety and phobias.1

"Models of stimulus generalization can be classified according to their "level of analysis. Among the models that address computational and algorithmic levels, two prominent types are 'feature-based' and 'rule-based' (Shanks and Darby 1998). Roughly speaking, feature-based models treat generalization as a bottom-up function of similarity in a feature-representation space, whereas rule-based models envisage generalization as a top-down process, based on abstract rules or categories that are either innate ("a priori) or acquired. Models that address the level of neuronal, 'hardware' implementation of generalization, date back to the early part of the twentieth century. Pavlov proposed that the neuronal basis of generalization is a spread of excitation from one specific brain region to another (Pavlov 1928). The idea was hence that the original stimulus was highly differentiated but the specificity lost due to the rich interconnectivity of the brain. The same type of concept could be applied to response generalization. This contrasts with a later hypothesis, that generalization is the original state and differentiation develops with experience (Lashley and Wade 1946). It is likely that in the brain both processes occur.

The circuit mechanisms of generalization are expected to depend on the type of task. Whenever cognitive generalization is concerned, the "hippocampal-cortical axis immediately comes to mind (McClelland and Goddard 1996). Which cellular and "synaptic mechanisms subserve generalization in these circuits? Is it at all subserved by distinct synaptic properties? Recent observations on "long-term potentiation (LTP) and synaptic specificity suggest candidate processes and mechanisms. LTP is considered as an input-specific mechanism that modifies only use-activated synapses. Engert and Bonhoeffer (1997) reported, however, that this specificity is rather limited: in their preparation, synapses at a distance of up to 70 |lm from the focus of potentiation were also potentiated, regardless of their activation history. This is a distance that can accommodate many synapses. Although the contribution of such neighbouring synapses to the relevant "representation is not yet known, this observation hints at a type of local process that might contribute to generalization. In another type of experiment, Frey and Morris (1997) found that the persistence of potentiation over time depends not only on local events at the activated synapse but also on prior activity of other synapses on the same neuron. Weak tetanic stimulation, which ordinarily leads to short-lived LTP only, resulted in long-term LTP, provided repeated tetanization had already been applied as long as 2-3 hours earlier at another input to the same population of neurons. This may be construed as yet another mechanism in which a specific synaptic input expands its sphere of effectiveness to nonactivated synapses ("context; for related findings in *Aplysia, see Martin et al. 1997a). Nevertheless, at present, the possibility that the aforementioned observations at the synaptic level contribute to generalization at the behavioural level should be treated only as provocative speculation.

Selected associations: Cue, Skill, Stimulus, Taxonomy, Transfer

1A refreshing example of overgeneralization is cited by Bandura and Walters (1963). This is a letter from the advice column of a leading metropolitan newspaper. And so it goes:

Dear Abby:

My girl friend fixed me up with a blind date and I should have known the minute he showed up in a bow tie that he couldn't be trusted. I fell for him like a rock. He got me to love him on purpose and then lied to me and cheated on me. Every time I go with a man who wears a bow tie, the same thing happens. I think girls should be warned about men who wear them. Against Bow Ties. Dear Against:

Don't condemn all men who wear bow ties because of your experience. I know many a man behind a bow tie who can be trusted.

Emergency Panic Remedies

Emergency Panic Remedies

Eliminate Your Panic And Anxiety Attacks Instantly With These Proven Techniques! There’s
 very
 little
 in
 this 
world
 that 
is
 as 
frightening
 and
 uncomfortable
 as 
a panic
 attack. 

Once
 they
 start,
 panic
 attacks
 can
 send
 you
 spiraling
 out
 of control.

Get My Free Ebook


Post a comment