Learning theories (child development)

 

Introduction

The normal human transition in cognitive ability from birth to maturity is vast and, as yet, without adequate explanation. During the last century, scientific consensus changed from a view that newborns possess virtually no knowledge of the world or of themselves to a view that they actually possess considerable innate bias that guides their interaction with their physical and social environment. The conception of learning mechanisms that might help explain the dramatic development of competence has undergone considerable change as well.

The idea that learning might play a strong role in development from birth to maturity has existed since the earliest of written history. Herodotus (485-425 BP) recorded in 440 BP concern for the role of learning in children’s development of language. The primary alternative to learning is usually stated in terms of the maturation of innate ability. While these opposing notions are of ancient origin, they continue their competition in developmental psychology at the present time.

In the past centuries, academic theories of learning were the outgrowth of scientists trying to replace folklore about the experiential role of repetition, effort, and temporal association in the production of such things as abilities, habits, and resilience of memories. The effort was to find objective laws like those being uncovered in physics and chemistry. For psychology, this would mean finding the laws that controlled the way experience changed an individual’s capacity and/or propensity to behave.

S-S learning

In 1927, when Ivan Petrovich Pavlov (1849-1936) published his classic work on the ‘conditioned reflex,’ it was seen to be exactly what had been hoped for. The loose notion of experience was replaced with defined categories of stimulus (S) and response (R), and further sub-categorical distinctions relating to the learning process, termed ‘conditioning.’ Thus, a law of learning was provided whereby systematic manipulation of stimuli would lead to a predictable change in an individual’s propensity to produce a particular response under specified conditions. Adding to the majesty of Pavlov’s laws of learning was the fact that timing of the manipulation of stimuli was very important, just as is so in the laws of physics and chemistry.

In the simplest terms, Pavlov uncovered a law for giving power to a stimulus that previously had little or none. The change in power was evidenced by the change in the conditioned individual’s response to the conditioned stimulus (CS). For example, as illustrated in Table 1, prior to applying the law of conditioning to a naive dog, the stimulus sound of a bell has no power to elicit salivation. However, after a number of pairings of this stimulus (the CS) just shortly prior to the presentation of meat (an unconditioned stimulus, UCS, that already possessed the power to cause salivation), the bell acquires the power to elicit salivation.

Pavlov’s experiment demonstrates how animals are sensitive to the temporal contingency between events. The dog experienced a series of trials in which presentation of meat was contingent on the bell having just sounded. The conditioning is dramatically weakened if the temporal contingency is reversed so that the meat precedes the bell. Pavlov investigated the lawful effects of varying the trial conditions and varying the timing between stimuli. Experiments of this kind are commonly referred to as ‘S-S learning’ and they continue to be investigated in many labs to this day. Renewed interest in this type of learning in humans has recently occurred in conjunction with techniques of brain imaging in studies of brain structure and function. For example, monitoring brain activity with event-related functional magnetic resonance imaging (efMRI) during classical conditioning of an angry face (CS) with an aversive sound (UCS) supports speculation of the special involvement of the amygdala in this form of learning (Morris, Buchel, & Dolan, 2001).

Table 1. S-S contingency learning: passing pre-existing power by classical conditioning.

Time Learning progression

t1 Sound of bell has no power to cause dog’s salivation.

Meat has power to cause dog’s salivation.

t2 Dog exposed to meat contingent on prior sound of bell for some number of conditioning trials.

t3 Sound of bell now has power to cause salivation.

John Broadus Watson (1878-1958) was probably the most eloquent promulgator of Pavlov’s laws as these might apply to development, particularly human development. He appeared to believe that S-S learning could account for virtually all of the variation in human ability and self-control. It was in relation to the latter, and his conditioning and extinguishing of fear in infants, that he gained a broad non-academic following for practical advice on child rearing during the 1920s and 1930s. Watson viewed his successful conditioning of fear as a blueprint for properly understanding the origins of the many irrational fears and emotional maladjustments that were often being treated by psychoanalysis at that time. An experiment by one of his students, Mary Cover Jones (1896-1987), extinguishing fear in an infant by presenting the fearful CS in association with naturally pleasant stimulation (UCS), has been acknowledged as the first experiment in therapeutic behavior modification, a clinical field begun in the 1950s with extensions of Pavlov’s findings by Joseph Wolpe (1915-1997).

Watson’s influence on the academic field of psychology and its effort to account for human development is arguably as great as and longer lived than his popular writings on child rearing. Pavlov’s laws of conditioning were perfect examples of the power of objectifying variables in the laws of psychology. Philosophically, Watson had been championing the virtue of behaviorism as a replacement for the subjective introspective methodology that psychology had relied on in its first steps as a science in the 19th century. One could, he argued, remove any reference to an individual’s subjective mental state when constructing laws that would adequately explain human behavior and development. Stimuli associated in specifiable patterns of temporal contingency would provide a sufficient causal reference in a philosophically sound science of psychology.

R-S and S-(R-S) learning

In the mid-thirties, Burrhus Fredric Skinner (19041990) introduced a behavioristic formalization of what was commonly viewed as goal-oriented or instrumental learning. It was a refinement and further objectification of what Edward Lee Thorndike (1874-1949) had termed the ‘law of effect.’ In contrast to Pavlov’s focus on the contingency between experiencing CS and UCS, Skinner’s focus was on the contingency between an unconditioned behavior, termed an operant response (R), and a subsequent unconditioned stimulus termed a reinforcer (Sre). He reframed the everyday notions of working for incentives and goals from their reference to anticipated future events which he found unacceptable. Skinner proposed instead a reliance solely on past contingency between behaviors and reinforcers that had followed them and stimuli that had marked these occasions of contingency. He believed that these basic categories of experience are sufficient to explain the development of even that most complex of human behaviors, language.

Table 2. S-(R-S) contingency learning: making new power by discriminative operant conditioning.

Time Learning progression

t1 Sound of bell has no power to cause dog to sit. Food has no power to cause dog to sit.

t2 Dog receives food contingent on sitting within five seconds of sound of bell for some number of learning trials.

t3 Sound of bell now has power to cause dog to sit.

We noted that S-S learning might be viewed as a method of passing power from one stimulus to another in terms of the power to elicit some specific behavior. In Pavlovian learning, however, the power that can be given to an initially powerless stimulus is limited to the set of available powers existing in unconditioned stimuli. In R-S learning theory, by contrast, power can be constructed in the absence of an available UCS for the target behavior. How this is done is termed discriminative learning (S-[R-S]). As summarized in Table 2, when the target behavior occurs, it is reinforced. This R-S contingency is itself made contingent on the presence of another stimulus. This new stimulus is called a discriminative stimulus (Sd) because it provides a basis to discriminate occasions in which the target behavior has been reinforced. After some number of trials in which the Sd is present and others in which it is absent, the individual will begin to emit the target behavior in response to the occurrence of the Sd. Thus, the trials have constructed an eliciting power for the Sd without a need for a UCS that possessed such eliciting power beforehand.

This expansion of learning theory to provide a behavioristic account of what is commonly called purposive or goal-oriented behavior was joined by a number of other notable theorists. Clark Hull (1884-1952) developed a set of inter-related lawful formulations that, like the laws of mechanics in physics, were intended to account for not only the occurrence but also the strength of learned behavior. Unlike Skinner, Hull and his student Kenneth Wartinbee Spence (1907-1967) tried to account for the power of some stimuli to function as rewards (or reinforcers in Skinner’s terminology) by appealing to a notion of biological needs that set up motivating ‘drives’ in the individual. Edward Chace Tolman (1886-1959) offered another significant variant of learning theory that explicitly tried to account for the philosophical notion that purposive behavior involves an individual possessing some cognitive representation of the goal being pursued. This shift from radical behaviorism to what was called purposive behaviorism opened the door for more recent theorizing about how maturation of memory and information processing capacities might alter learning and its effect on subsequent development.

Although Skinner did not receive the degree of attention that Watson did in the area of popular guidance for child rearing, Skinner’s theoretical perspective has had deep and lasting influence in the USA in the areas of special education and classroom management of children in primary school grades. This may be due to the relative clarity and simplicity of his prescriptions for modifying behavior. It may also be due in part to Skinner’s novel Walden Two (1948), which was, in effect, a treatise on his view of ideal rearing conditions. That topic was widely read in undergraduate courses of humanities and social philosophy during the second half of the twentieth century.

The radical behaviorism that Watson and Skinner espoused greatly influenced academic psychology in the United States until the mid 1950s. This was less true in Europe and within the sub-field of developmental psychology where Freudian psychoanalytic theory, the cognitive developmental theory of Jean Piaget (1896-1980), and various bio-maturational theories maintained many adherents regarding how maturation and early experience affect human development. In the USA, there was a productive tension in the lingering theoretical struggle between learning theory and psychoanalytic theory that resulted in the collaborative collection and analysis of cross-cultural data in an archive known as the Human Relations Area Files (HRAF) at Yale University (now accessible on the Internet). Initially, cultural anthropologists and developmental psychologists compiled observations of the variation in child rearing as this could be discerned across a variety of reasonably independent cultures (now more than ninety) accounted for by the HRAF. The ethnographic observations made were in part guided by the goal of testing differences in the developmental predictions that would follow from the theoretical frameworks of learning theory and Freudian theory. Overall, learning theory has tended to fair better in this contest.

Social learning theory

Albert Bandura (1925- ) and others loosened the grip of behaviorism on learning theory. This occurred in three important respects. Initially, it was on the basis of a specific concern for the process of imitation or observational learning. Bandura and his colleagues pointed out that when a child changed a propensity to behave in a certain way simply by observing another person perform that behavior, the learning (i.e., the change in propensity) should be viewed as occurring in the absence of a learning trial. That is, imitation appeared to be unlike the simple laws of Pavlovian S-S learning or Skinnerian R-S learning, wherein the conditioning of the individual involved some number of trials in which the focal behavior occurred – as elicited by the UCS in the case of Pavlovian learning or as emitted and reinforced in the case of Skinnerian learning. By contrast, imitation involved only the observation of another individual enacting the target behavior. Bandura highlighted the uniqueness of this by labeling it ‘no-trial learning.’ In addition, the fact that the ease and strength of imitation was found to vary in relation to social characteristics of the model was a serious barrier to any concerted effort to classify stimuli in strictly physical terms as was the preference of radical behaviorism.

The third issue raised by Bandura’s work, and that of others, was the effect of the individual’s attribution of control to the model. Imitation is more likely if the observer evaluates the model’s behavior as truly causing the apparent consequences. However, the research of Andrew Meltzoff with newborns indicates that imitation can proceed without such contextual evaluation in the beginning. On the other hand, a recent study suggests that imitation by 14-month-olds depends in part on whether the model’s behavior is perceived as rational under the circumstances in which the behavior is modeled (Gergely, Bekkering, & Kiraly, 2002). The preceding findings and related work with animals (e.g., Michael Tomasello’s work with chimpanzees) has lead to debate regarding the possibility that imitation develops from a response with no inference, to a response based on a rational and eventually an intentional stance. Some extension of this debate may come from brain imaging research such as that of Jean Decety and his colleagues who have recently used positron emission tomography (PET) to distinguish the neural mechanisms underlying acts of imitation and the perception of being imitated.

Table 3. Four ways to perceive contingency of event E2 on event El.  


Basis of perception

Mechanism involved

Contiguity

Detection of short time span between instances of E1 and E2

Correlation

Computation of co-variation in time of instances of E1 and E2

Conditional probability

Computation of probability of E2 in time following instances of E1 and probability of E1 in time preceding instances of E2

Logical implication

Deduction of contingency by combining evidence of truth for each of the following relations: E1 implies E2, E1 implies not-E2, not-E1 implies E2, and not-E1 implies not-E2

New forms of learning theory

The learning process, as conceived by the radical behaviorists, was meant to stand objectively independent of the learner. We have noted above that later theorizing has seriously eroded the independence of learning from the subjective/cognitive processes of the learner. The effectiveness of a contingency in S-S learning, R-S learning, or even in learning by imitation, now appears to become soon dependent on the cognitive-perceptual activity of the individual. Learning depends on contingency, but on contingency as perceived by the learner. When put in this perspective, it becomes important to consider just how a contingency (be it between stimuli or between behavior and stimuli) might be perceived.

Contingency perception

As outlined in Table 3, there have been at least four theoretical proposals for how contingency might be perceived: contiguity, correlation, conditional probability, and logic. Contiguity was favored by the radical behaviorists. It had the appeal of simple dependence on temporal separation. The greater the time between events, the less likely the learning. The generality of this ‘law of contiguity was seriously undermined in the 1960s by John Garcia who found food aversion learning in which strong S-S learning occurred despite extensive delay between a novel food CS and a subsequent noxious UCS. It was further undermined by findings of’learned helplessness’ wherein R-S learning fails to occur despite short delay between behavior and normally reinforcing stimuli that previously were experienced as being non-contingent or independent of behavior (Peterson, Maier, & Seligman, 1993).

The proposals of contingency being perceived in terms of correlation, conditional probability, or logic each carry a computational assumption regarding the learner’s capacity to evaluate the experience of a contingency. These three potential indices of contingency respectively make reference to progressively more details in the representation of the contingency. Correlation centers on a single index of co-variation between, say, events E1 and E2. Conditional probability introduces two indices, the prospective probability of E2 occurring given E1 has occurred, and the retrospective probability of E1 having occurred given E2 occurs. Logical inference, as proposed by Thomas G. R. Bower, involves evidence in relation to four possible connections: E1 implies E2, E1 implies not-E2, not-E1 implies E2, and not-E1 implies not-E2. Most work to date has been framed in terms of either contiguity or conditional probability.

Constraints on learning

In recent decades, as learning theory has become more cognitive, it has also become increasingly integrated with evolutionary theory. The earlier hopes of finding universal laws of learning that would account for development similarly across species were strained by the reports of species-specific imprinting uncovered by ethologists as well as by the species variation of food aversion learning reported by Garcia and others. Likewise, the attempt to encompass the complexity of language acquisition in laws of R-S contingencies by Skinner and others lost favor to the seemingly more adequate account provided by Noam Chomsky and his students that incorporated an assumption that humans are innately equipped with an abstract learning mechanism containing the primitive categories of language structure. Language learning was thus a very constrained matter of discriminating sound patterns that fit the preset linguistic categories. Of note was the apparent fact that language learning progressed without the need for reward and in a manner that seemed self-guided, as evidenced in such phenomena as the English-speaking child’s errors of over-regularization of the past tense suffix ‘-ed’ for previously mastered cases of irregular verbs (e.g., “mommy goed out”). Such special constraints on general learning laws suggest that learning mechanisms are part of the evolved equipment each species has obtained in its adaptation to environmental pressure.

John Bowlby (1907-1990) was influenced by the findings of ethology and the growth of cognitive learning theory in his construction of a new theory of early socio-emotional development called attachment theory. Bowlby replaced Freud’s speculations (basically Pavlovian in form) about the role of stimulus associations in children’s development of emotional attachment to their parents. He emphasized instead the interplay of proximity control signals that have evolved in humans. In this view, children clearly must learn who to love and how to manage their emotional states. However, rather than building this learning through associations of people and stimulus events that reduce somatic tension as claimed by Freud, Bowlby proposed a central role for a restricted sub-set of stimulus events that were part of an evolved system of infant and parental behaviors that would help them maintain spatial proximity. The system served to protect immature members of our species from predation, especially in the prehistoric environments of our ancestors or what Bowlby called the “environment of evolutionary adaptedness.”

Contingency as a stimulus

The idea that the contingencies of learning need to be perceived in order to be effective undermines the prospect of finding laws of learning that can be formulated without concern for the mental processes of the learner. It also introduces a new option for the potential effects of contingency experience. As an object of perception, contingency experience can be viewed as a stimulus in its own right. From this perspective, the infant controlling the mobile in Figure 1 by movement on pressure-sensitive pillows is not only perceiving the mobile turn, but the contingency of its turning as well. A responsive mother can display her attention to her baby by any number of auditory, tactile, or visual reactions to any number of actions on the part of her infant. The baby’s perception that her reactions are contingent on his behavior may have a psychological impact that is separate from and far more important than the stimulus impact of her behavior itself. In this manner, contingency experience has been proposed to have unconditioned eliciting power in human infants in their early phases of social development. In this view, part of what initially defines the mother is her characteristic level of contingency as perceived by the infant. Moreover, this view allows that contingency may be misperceived and certain forms of misperception may have predictable developmental consequences (Watson, 2001).

Infant learning to control mobile movement using pressure-sensitive pillows.

Figure 1. Infant learning to control mobile movement using pressure-sensitive pillows.

Probability learning

Studies examining the sensitivity of infants and young children to the contingency structure of their environment have increased in recent years. These range from testing how contingent responsiveness of a mobile affects subsequent reactions to it to how inter-stimulus contingency level may be used to segregate words in speech. The latter work is an interesting challenge to one aspect of Chomsky’s proposed abstract linguistic categories.

Chomsky and his students have argued that linguistic structure had such variable distribution in the continuous stream of natural speech that it was not decipherable without some prior knowledge. This argument for an inborn guidance to language development (or so-called ‘language acquisition device’) has been weakened by the discovery that the transitional probability structure of the phonemic sequences in natural speech can reveal at least some of its underlying linguistic structure. Moreover, researchers have shown that 8-month-old infants are sensitive to conditional probability structure in continuous streams of an artificial language (Aslin, Saffran, & Newport, 1998). Thus, it would seem that some of the presumably innate grounding of language development may really be provided by the human infant’s capacity for probability learning.

Neural-net models and new views of learning and maturation

Since the early 1990s, the elaboration of cognitive learning models has been spurred on by techniques of computer simulation of neural-net learning. Despite the seminal work of Donald O. Hebb (1904-1985), the most influential learning theories of the past century did not speculate as to how their learning laws were supported by an animal’s neurophysiological structure. The picture is quite the opposite today. Neural-net modeling, for example, is an explicit attempt to approximate the mechanics of neurological adaptation to experience. These models have just primitive features, such as neuronal nodes, their activation, their synaptic-like inter-connection, their inter-node conductivity (termed ‘weights’), and their form of inter-connection across layers of nodes. The neural-net simulations have been applied to a wide variety of classic developmental issues. Although neural nets embody no use of symbolic representation, they have managed to learn to perform tasks that had previously been assumed to require symbolic thought. For example, James L. McClelland and his colleagues have shown that neural nets can do quite well in simulating the developmental progression in solving various conservation tasks that were highlighted in the classic work on human cognitive development by Piaget.

Neural-net modeling of learning has spawned ideas about how maturation and learning maybe viewed as cooperative in the explanation of development. Jeffrey L. Elman found that a neural net was incapable of mastering an artificial language even though the net had a theoretically sufficient amount of memory capacity as provided by feedback connections between layers. The net could master the language if the linguistic examples were initially simple and then later of full complexity. However, real world learning does not occur with sequential exposure to partial then full language structure. So Elman arranged to have the net begin with a limited memory followed by a maturational shift to full memory. In this case, the net succeeded (Elman, 1993). It seems reasonable to expect that evolution may have worked out a cooperation between maturation and learning for many recurrent environmental challenges.

Table 4. Potential developmental consequences of contingency experience.  

S-S contingencies


R-S and S-(R-S) contingencies

Extend range of stimuli that cause behavioral reactions (e.g., learn to fear fire).

Increase strength and likelihood of behavioral reactions and create new causal classes of stimuli for discriminative control of those behaviors (e.g., learn to squeeze toy to hear squeak).

Learn categories of S-S contingency structure in sequential stimuli (e.g., learn the word segmentation of linguistic utterances).

Learn categories of responsive objects (e.g., learn mom is person who is especially responsive to cry).

Learn causal powers of environmental objects (e.g., learn that only certain objects will float).

Learn extent of personal efficacy (e.g., learn where, when, and what things can be caused by one’s own behavior).

Learn association of feelings and sights of own action (e.g., learn hand motion sequence to assist imitation of mom’s clapping of hands).

Learn social signals that mark change in responsiveness of others (e.g., learn the face and body confirmations that display the mood, emotional states, and intentions of another person)

Conclusions

The past century has seen a dramatic shift in theorizing about the role of learning in development. Dominance has shifted from a preference for formulating laws of contiguity between external stimuli and behavior to a preference for formulating models with various degrees of biologically based computational activity and novel assumptions of evolved bias in guiding adaptation to the contingency structure of the physical and social environment (Table 4). In addition, recent advances in brain imaging and neural-net simulation techniques hold some promise for further insights into the biological substrate of learning mechanisms and their evolved constraints.

Next post:

Previous post: