Previous laboratory studies on social learning suggest that some animals can learn more readily if they observe a conspecific demonstrator perform the task unsuccessfully and so fail to obtain a food reward than if they observe a successful demonstrator that obtains the food. This effect may indicate a difference in how easily animals are able to associate different outcomes with the conspecific or could simply be the result of having food present in only some of the demonstrations. To investigate we tested a scatter-hoarding mammal, the eastern grey squirrel, on its ability to learn to choose between two pots of food after watching a conspecific remove a nut from one of them on every trial. Squirrels that were rewarded for choosing the opposite pot to the conspecific chose correctly more frequently than squirrels rewarded for choosing the same pot (a feature-negative effect). Another group of squirrels was tested on their ability to choose between the two pots when the rewarded option was indicated by a piece of card. This time, squirrels showed no significant difference in their ability to learn to choose the same or the opposite pot. The results add to anecdotal reports that grey squirrels can learn by observing a conspecific and suggest that even when all subjects are provided with demonstrations with the same content, not all learning occurs equally. Prior experience or expectations of the association between a cue (a conspecific) and food influences what can be learned through observation whilst previously unfamiliar cues (the card) can be associated more readily with any outcome.

ISSN: 0963-7214

BBSRC Grant 9/S17109
EC Framework 6 project grant 516542 (NEST)

Many formal models of categorization assume, implicitly or explicitly, that categorization results in the formation of direct associations from representations of the presented stimuli to representations of the experimentally-provided category labels. In three categorization experiments employing a polymorphous classification structure (Dennis, Hampton, & Lea, 1973) and a partial reversal, “optional shift” procedure, (Kendler, Kendler, & Wells, 1960) we provide evidence consistent with the hypothesis that learning a new classification problem results in the creation of category representations that mediate between representations of stimulus and label. This hypothesis can be instantiated through the AMBRY model (Kruschke, 1996).

Two experiments examined the outcome-specificity of a learned predictiveness effect in human causal learning. Experiment 1 indicated that prior experience of a cue–outcome relationship modulates learning about that cue with respect to a different outcome from the same affective class, but not to an outcome from a different affective class. Experiment 2 ruled out an interpretation of this effect in terms of context-specificity. These results indicate that learned predictiveness effects in human causal learning index an associability that is specific to a particular class of outcomes. Moreover they mirror demonstrations of the reinforcer-specificity of analogous effects in animal conditioning, supporting the suggestion that, under some circumstances, human causal learning and animal conditioning reflect the operation of common associative mechanisms.

    In April 1903, Pavlov presented his early work on the conditioned reflex to the International Congress of Medicine, Madrid (Pavlov, 1903/1928a). Two of the, no doubt numerous, events that celebrated the centenary of this event were the Spring meeting of the Experimental Psychology Society at Exeter (for which I was the local organizer) and the annual Associative Learning Symposium organized by Professors Rob Honey and John Pearce of Cardiff University.
    As the date of the Experimental Psychology Society (EPS) conference approached, my thoughts turned increasingly to the nature of the man and the work we were about to celebrate, and how both were considered today. I repeatedly recalled a conversation with a colleague about a year earlier that started with him saying, 'Associative learning ... isn't that something they did in the 1950s when they thought pigeons could be trained to guide missiles?"
    It's probably best that this particular opinion remains anonymous but, in less extreme forms, it probably reflects the attitude of a considerable minority of today's professional psychologists. Breaking down the opinion into its constituents, it seems to center on the ideas that the study of associative learning is (a) entirely concerned with nonhuman animals, and is (b) part of an episode in the history of psychology that ended with the "cognitive revolution" of the 1960s and 1970s. Both components of this opinion are fundamentally incorrect, as I indicate later in this chapter. However, first there is a need to define terms....

In this chapter, we describe three basic associative theories of learning as they might be applied to the problem of category learning. First, we describe Hebbian learning, which was one of the earliest formal associative theories of learning. Second, we describe competitive learning, which can, in some ways, been seen as a development of Hebbian learning, and is a theory of unsupervised learning. next, we describe the Rescorla-Wagner theory, which is a theory of supervised learning. We then argue that there is a need for a mechanism that can use feedback when it is available (supervised learning) but will continue to learn in feedback is absent (unsupervised learning). We propose a possible mechanism that involves adding certain aspects of Rescorla-Wagner theory to competitive learning. Our proposed system makes a claer prediction about people's behavior in a free-classification task, and we describe how we have begun to test that prediction. The chapter ends with a consideration of other theoretical approaches to the problem of category learning, both within the domain of associative learning and more generally.

What are the effects of pre-exposure of stimuli on participants' subsequent ability to categorize them accurately? An experiment employing artificial, abstract, visual stimuli confirms that, for adult humans, the effect of pre-exposure is dependent upon category structure. Whether pre-exposure has beneficial or detrimental effects is shown to be dependent on the way category examples are generated from the category base patterns. The results are predicted by salience reduction accounts of perceptual learning but may be problematic for stimulus differentiation accounts.

It has been demonstrated that when individuals free classify stimuli presented simultaneously in an array, they have a preference to categorize by a single dimension (e.g. Medin, Wattenmaker & Hampson, 1987). However, when people are encouraged to categorize items sequentially, initial research has suggested that people sort by “family resemblance” – grouping by overall similarity (Regehr &  Brooks, 1995). Our studies extended this research, producing three main findings. First, the sequential procedure introduced by Regehr & Brooks (1995) does not always produce a preference for family resemblance sorts. Second, sort strategy in a sequential procedure is highly sensitive to subtle variations in stimulus properties. Third, and surprisingly in the context of previous research, spatially separable stimuli evoked more family resemblance sorts than stimuli of greater spatial integration. It is suggested that the family resemblance sorting observed is the result of an analytic strategy as opposed to a perceptually driven non-analytic strategy.

When humans process visual stimuli, global information often takes precedence over local information. In contrast, some recent studies have pointed to a local precedence effect in both pigeons and nonhuman primates. In the experiment reported here, we compared the speed of acquisition of two different categorizations of the same four geometric figures. One categorization was on the basis of a local feature, the other on the basis of a readily apparent global feature. For both humans and pigeons, the global-feature categorization was acquired more rapidly. This result reinforces the conclusion that local information does not always take precedence over global information in nonhuman animals.

Artificial creatures form an increasingly important component of interactive computer games. Examples of such creatures exist which can interact with each other and the game player and learn from their experiences. However, we argue, the design of the underlying architecture and algorithms has to a large extent overlooked knowledge from psychology and cognitive sciences. We explore the integration of observations from studies of motivational systems and emotional behaviour into the design of artificial creatures. An initial implementation of our ideas using the “sim agent” toolkit illustrates that physiological models can be used as the basis for creatures with animal like behaviour attributes. The current aim of this research is to increase the “realism” of artificial creatures in interactive game-play, but it may have wider implications for the development of AI.

The effects of limited processing time were investigated for the binary categorization of artificial multidimensional objects. Following Lamberts and Freeman (1999), in the first stage of the experiment participants learnt to categorize 9 stimuli into two categories. In the second stage, the same stimuli were presented for categorization, and both the display time, and the time available in which to make a decision, were varied independently. It was found that each of these variables had a significant effect on accuracy of categorization, as well as response latency. Lamberts and Freeman (1999) demonstrated that restricting presentation time of a certain stimulus in their category structure caused a reversal in category assignment. We found evidence of the same reversal, but it was dependent on the time available to make a decision rather than the duration of stimulus display. Importantly, changes in accuracy due to response deadline were not explicable in terms of truncation of processing by the limited time. The study provides an empirical investigation of the intuitive notion that both perceptual processing and decision making components are time dependent.

Rumelhart & Zipser's (1986) competitive learning algorithm is an account of unsupcrvised learning and, as such, might be considered a potential model of free classification behavior in humans. However, selective learning effects (e.g. Dickinson, Shanks & Evenden, 1984) suggest that human learning, ai least under conditions of feedback, may be better characterized by an error-correcting system. An experiment is reported that provides preliminary evidence for the existence of a selective learning effect in free classification. Simulations indicate that Rumelhart & Zipser's algorithm does not provide an adequate account of the behavior observed, whilst an error-correcting variant of competitive learning does.

The effect of a response deadline on categorical decisions was investigated. Time available for response was manipulated in the test phase, along with stimulus difficulty. Effects of these manipulations were observed in response accuracy, and in the mean, standard deviation and skew of the reaction times. The effects observed demonstrate that participants responded to the deadline in an adaptive manner - reducing their reaction time to long-latency decisions whilst leaving short latency decisions relatively unaffected. A simple connectionist model of categorical decisions (Wills & McLaren, 1997) is shown to account for this behavior.

Many theories of learning and memory (e.g., connectionist, associative, rational, exemplar based) produce psychological magnitude terms as output (i.e., numbers representing the momentary level of some subjective property). Many theories assume that these numbers may be translated into choice probabilities via the ratio rule, also known as the choice axiom (Luce, 1959) or the constant-ratio rule (Clarke, 1957).We present two categorization experimentsemploying artificial, visual, prototype-structured stimuli constructed from12 symbols positioned on a grid. The ratio rule is shown to be incorrect for these experiments, given the assumption that the magnitude terms for each category are univariate functions of the number of category-appropriate symbols contained in the presented stimulus. A connectionist winner-take-all model of categorical decision (Wills & McLaren, 1997) is shown to account for our data given the same assumption. The central feature underlying the success of this model is the assumption that categorical decisions are based on a Thurstonian choice process (Thurstone, 1927, Case V) whose noise distribution is not double exponential in form.

Although it is currently popular to model human associative learning using connectionist networks, the mechanism by which their output activations are converted to probabilities of response has received relatively little attention. Several possible models of this decision process are considered here, including a simple ratio rule, a simple difference rule, their exponential versions, and a winner-take-all network. Two categorization experiments that attempt to dissociate these models are reported. Analogues of the experiments were presented to a single-layer, feed-forward, delta-rule network. Only the exponential ratio rule and the winner-take-all architecture, acting on the networks’ output activations that corresponded to responses available on test, were capable of fully predicting the mean response results. In addition, unlike the exponential ratio rule, the winner-take-all model has the potential to predict latencies. Further studies will be required to determine whether latencies produced under more stringent conditions conform to the model’ s predictions.

Two experiments are reported that investigate the difference in g radient of generalization observed between one-category (non-discriminative) and two-category (discriminative) training. Extrapolating from the results of a number of animal lear ning studies, it was predicted that the gradient should be steeper under discriminative tr aining. The first experiment concerns this basic prediction for the stimuli used, which were novel, prototype-structured, and constructed from 12 symbols positioned on a grid. An explanation for the effect, based on the Rescorla-Wagner theory of Pavlovian conditioning (Rescorla & Wagner, 1972), is that under non-discriminative tr aining "incidental stimuli" have significant control over responding, whereas under discriminative training they do not. Incidental stimuli are those aspects of the stimulus, or the surrounding context, that are not differentially reinforced under discriminative training. This explanation leads to the prediction that a comparable effect of blocked ver sus intermixed discriminative training should also be found. This prediction is disconfirmed by the second experiment. An alternative model, still based on the Rescorla-Wagner theory but with the addition of a decision mechanism comprising a threshold unit and a competitive network system, is proposed, and its ability to predict both the choice probabilities and the pattern of response times found is evaluated via simulation.