Population of linear experts: knowledge partitioning and function learning

Knowledge partitioning is a theoretical construct holding that knowledge is not always integrated and homogeneous but may be separated into independent parcels containing mutually contradictory information. Knowledge partitioning has been observed in research on expertise, categorization, and function learning. This article presents a theory of function learning (the population of linear experts model--POLE) that assumes people partition their knowledge whenever they are presented with a complex task. The authors show that POLE is a general model of function learning that accommodates both benchmark results and recent data on knowledge partitioning. POLE also makes the counterintuitive prediction that a person's distribution of responses to repeated test stimuli should be multimodal. The authors report 3 experiments that support this prediction.     

Context-gated knowledge partitioning in categorization

According to the knowledge partitioning framework, people sometimes master complex tasks by creating multiple independent parcels of partial knowledge. Research has shown that knowledge parcels may contain mutually contradictory information, and that each parcel may be used without regard to knowledge that is demonstrably present in other parcels. This article reports 4 experiments that investigated knowledge partitioning in categorization. When component boundaries of a complex categorization were identified by a context cue, a significant proportion of participants learned partial and independent categorization strategies that were chosen on the basis of context. For those participants, a strategy used in one context was unaffected by knowledge demonstrably present in other contexts, suggesting that knowledge partitioning in categorization can be complete.     

Knowledge partitioning in categorization: Boundary conditions

Knowledge partitioning refers to the notion that knowledge can be held in independent and nonoverlapping parcels. Partitioned knowledge may cause people to make contradictory decisions for identical problems in different circumstances. We report two experiments that explored the boundary conditions of knowledge partitioning in categorization. The studies examined whether or not people would partition their knowledge (1) when categorization rules were or were not verbalizable and (2) when the to-be-categorized stimuli comprised perceptually separable or integral dimensions. When learning difficulty was controlled, partitioning occurred across all combinations of verbalizability and integrality/separability, underscoring the generality of knowledge partitioning. Partitioning was absent only when the task was rapidly learned and people reached a high level of proficiency, suggesting that task difficulty plays a critical role in the emergence of partitioned knowledge.     

Knowledge partitioning in categorization: constraints on exemplar models

The authors present 2 experiments that establish the presence of knowledge partitioning in perceptual categorization. Many participants learned to rely on a context cue, which did not predict category membership but identified partial boundaries, to gate independent partial categorization strategies. When participants partitioned their knowledge, a strategy used in 1 context was unaffected by knowledge demonstrably present in other contexts. An exemplar model, attentional learning covering map, was shown to be unable to accommodate knowledge partitioning. Instead, a mixture-of-experts model, attention to rules and instances in a unified model (ATRIUM), could handle the results. The success of ATRIUM resulted from its assumption that people memorize not only exemplars but also the way in which they are to be classified.     

Feature distribution and background knowledge in category learning

Three experiments examined the interactions of category structure and prior knowledge in category learning.Experiment 1 examined the distribution of atypical or 'crossover' features in category learning. In real categories, crossover features may be unevenly distributed found primarily in very unusual examples of a category (like whales or ostriches). In contrast, in many psychology experiments, each item has exactly one crossover feature. Even versus uneven distribution of crossover features did not affect category learning when the features were neutral. However, when the features were connected by prior knowledge, it was much harder for subjects to learn the structure with the uneven distribution of crossover features. Experiments 2 and 3 found similar results with a slightly less uneven condition. We conclude that learning is a function of the interaction of category structure and prior knowledge rather than either one alone. Furthermore, knowledge benefits learning even when the category contains contradictions of the knowledge, so long as the contradictions are not very salient.     

Landmarks as beacons and associative cues: Their role in route learning

Relatively little is known about how people use the landmarks in their environment to learn routes. Landmarks are commonly regarded as associative cues--stimuli that enable recall of directional responses that lead closer to the navigator's goal. We contrast the function of landmark as associative cue with that of a beacon-a landmark near enough to a goal that moving toward it leads the navigator closer to his or her goal. In five experiments, participants learned a route through a simple desktop virtual environment. In the first three experiments, routes were learned better when their landmarks served as beacons than as associative cues. Two additional experiments showed that the acquired route knowledge depends on the function that landmarks serve during learning. Beacon-based route knowledge is less enduring and relatively less likely to involve knowledge of directions in the environment than is the route knowledge formed from landmarks that serve as associative cues.     

A knowledge-resonance (KRES) model of category learning

This article introduces a connectionist model of category learning that takes into account the prior knowledge that people bring to new learning situations. In contrast to connectionist learning models that assume a feedforward network and learn by the delta rule or backpropagation, this model, the knowledge-resonance model, or KRES, employs a recurrent network with bidirectional symmetric connection whose weights are updated according to a contrastive Hebbian learning rule. We demonstrate that when prior knowledge is represented in the network, KRES accounts for a considerable range of empirical results regarding the effects of prior knowledge on category learning, including (1) the accelerated learning that occurs in the presence of knowledge, (2) the better learning in the presence of knowledge of category features that are not related to prior knowledge, (3) the reinterpretation of features with ambiguous interpretations in light of error-corrective feedback, and (4) the unlearning of prior knowledge when that knowledge is inappropriate in the context of a particular category.     

Meet The Simpsons: top-down effects in face learning

We examined whether prior knowledge of a person affects the visual processes involved in learning a face. In two experiments, subjects were taught to associate human faces with characters they knew (from the TV show The Simpsons) or characters they did not (novel names). In each experiment, knowledge of the character predicted performance in a recognition memory test, relying only on old/new confidence ratings. In experiment 1, we established the technique and showed that there is a face-learning advantage for known people, even when face items are counterbalanced for familiarity across the experiment. In experiment 2 we replicated the effect in a setting which discouraged subjects from attending more to known than unknown people, and eliminated any visual association between face stimuli and a character from The Simpsons. We conclude that prior knowledge about a person can enhance learning of a new face.     

Insight and strategy in multiple-cue learning

In multiple-cue learning (also known as probabilistic category learning) people acquire information about cue-outcome relations and combine these into predictions or judgments. Previous researchers claimed that people can achieve high levels of performance without explicit knowledge of the task structure or insight into their own judgment policies. It has also been argued that people use a variety of suboptimal strategies to solve such tasks. In three experiments the authors reexamined these conclusions by introducing novel measures of task knowledge and self-insight and using "rolling regression" methods to analyze individual learning. Participants successfully learned a four-cue probabilistic environment and showed accurate knowledge of both the task structure and their own judgment processes. Learning analyses suggested that the apparent use of suboptimal strategies emerges from the incremental tracking of statistical contingencies in the environment.     

Context dependent learning in the serial RT task

This study investigated the development of contextual dependencies for sequential perceptual-motor learning on static features in the learning environment. In three experiments we assessed the effect of manipulating task irrelevant static context features in a serial reaction-time task. Experiment 1 demonstrated impaired performance after simultaneously changing display color, placeholder shape, and placeholder location. Experiment 2 showed that this effect was mainly caused by changing placeholder shape. Finally, Experiment 3 indicated that changing context affected both the application of sequence knowledge and the selection of individual responses. It is proposed either that incidental stimulus features are integrated with a global sequence representation, or that the changed context causes participants to strategically inhibit sequence skills.     

Re-examining the role of context in implicit sequence learning

Implicit sequence learning typically develops gradually, is often expressed quite rigidly, and is heavily reliant on contextual features. Recently we reported results pointing to the role of context-specific processes in the acquisition and expression of implicit sequence knowledge (D’Angelo, Milliken, Jiménez, & Lupiáñez, 2013). Here we examined further the role of context in learning of first-order conditional sequences, and whether context also plays a role in learning second-order conditional structures. Across five experiments we show that the role of context in first-order conditional sequences may not be as clear as we had previously reported, while at the same time we find evidence for the role of context in learning second-order conditional sequences. Together the results suggest that temporal context may be sufficient to learn complementary first-order conditional sequences, but that additional contextual information is necessary to concurrently learn higher-order sequential structures.     

Prior knowledge and correlational structure in unsupervised learning.

Prior knowledge has been shown to facilitate both supervised and unsupervised category learning, but questions remain about how this facilitation occurs. This article describes two experiments that investigate the effects of prior knowledge on unsupervised learning, using the exemplar-memory task of Clapper and Bower (2002). Experiment 1 demonstrates that prior knowledge facilitates learning in this task, as expected, and that this facilitation extends to both knowledge-relevant and knowledge-irrelevant features of the new categories. Experiment 2 shows that knowledge facilitates learning not only by increasing the probability that people will discover separate categories, but also by making the features of different categories seem less interchangeable, thereby reducing interference and confusion among them. Taken together, these experiments demonstrate that prior knowledge has multiple effects on unsupervised learning and suggests that the exemplar-memory task may provide a useful procedure for disentangling and investigating these effects.     

The effect of category learning on sensitivity to within-category correlations

A salient property of many categories is that they are not just sets of independent features but consist of clusters of correlated features. Although there is much evidence that people are sensitive to between-categories correlations, the evidence about within-category correlations is mixed. Two experiments tested whether the disparities might be due to different learning and test tasks. Subjects learned about categories either by classifying items or by inferring missing features of items. Their knowledge of the correlations was measured with classification, prediction, typicality, and production tests. The inference learners, but not the classification learners, showed sensitivity to the correlations, although different tests were differentially sensitive. These results reconcile some earlier disparities and provide a more complete understanding of people's sensitivities to within-category correlations.     

The influence of prior knowledge in intentional versus incidental concept learning

Previous research has demonstrated that background knowledge has a clear influence on concept learning. This influence, however, has been observed with a narrow range of intentional learning tasks. In the present experiments, the role of background knowledge was examined as a function of a variety of incidental learning tasks as well as with intentional learning tasks. The influence of prior knowledge was investigated by comparing the encoding of conceptually related co-occurrences with the encoding of conceptually unrelated co-occurrences. A clear influence of prior knowledge was observed with incidental encoding, and conceptual relatedness was found to have at least as powerful an influence with incidental as with intentional learning tasks. The results indicate that many types of knowledge-based influences will not vary as a function of encoding strategy. The pervasiveness and strength of the influence of background knowledge on concept learning are discussed.     

Role of context in perceptual learning in maze discriminations

Three experiments with rats in a maze examined the effects of pre-exposure to the relevant discriminative stimuli (rubber and sandpaper-covered maze arms) or the extra-maze context (the maze was surrounded either by black curtains or by variety of extra-maze landmarks) on the learning of a discrimination between rubber and sandpaper arms. In Experiment 1, pre-exposure to the extra-maze context facilitated subsequent discrimination learning. Experiments 2 and 3 showed that pre-exposure to rubber and sandpaper arms facilitated subsequent discrimination learning only when these cues were presented in the same context during pre-exposure and discriminative training. Taken together, the results are consistent with the hypothesis that a major cause of perceptual learning is the latent inhibition of stimuli or features common to the two discriminative stimuli, and that such latent inhibition may be disrupted by a radical change of context.     

The acquisition of category structure in unsupervised learning

Four experiments examined the extent to which prior knowledge influences the acquisition of category structure in unsupervised learning conditions. Prior knowledge is general knowledge about a broad domain that explains why an object has the features it does. Category structure refers to the statistical regularities of features within and across categories. Subjects viewed items and then divided them up into the categories that seemed most natural. Each item had one feature that was related to prior knowledge and five features that were not. The results showed that even this small amount of prior knowledge helped subjects to discover the category structure. In addition, prior knowledge enhanced the learning of many of the category's features, and not just the features that were directly relevant to the knowledge. The results suggest that prior knowledge may help to integrate the features of a category, thereby improving the acquisition of category structure.     

Effect of learning on production of duration in variable motor conditions

It has been suggested that increasing proprioceptive feedback and ensuring its consistency from trial to trial favours timing accuracy. The progressive acquisition of a timing performance with isometric and anisometric responses was investigated in conditions of 'inconsistent feedback', with two different methods: counting seconds or not. Subjects gripped the handle of a dynamometer during precisely 5 seconds in 4 randomly distributed conditions: 'Weak' or 'Strong' constant force, 'Slow' or 'Rapid' variable force. A first group produced the durations without counting and a second group counted seconds either mentally or aloud. Learning was examined in 4 successive blocks of trials. Average produced durations did not differ as a function of group, but the variability was reduced when subjects counted seconds. In both groups, the constant force conditions induced more accurate responses than the variable force conditions in the first block of trials. 'Slow' and 'Rapid' conditions were respectively associated with overestimation and under-estimation of response duration. These trends diminished progressively with learning. Both groups yielded sequential effects linked to duration and independent of condition. These data suggest that, whether subjects counted or not, learning was not based on condition-specific proprioceptive feedback. It may involve the elaboration of an internal temporal reference common to all conditions.     

Implicit learning of new verbal associations

Implicit learning of a series of new verbal associations was studied in four experiments. The first two experiments demonstrated that learning of a repeating sequence of verbal stimuli may occur without awareness, but only when the stimulus-response mapping requires an attention-demanding activity: Subjects who were unaware of the sequence learned when instructed to categorize the stimuli, but not when instructed simply to read them. However, in both situations, unaware subjects performed no better than untrained control subjects in expressing their knowledge of the sequence explicitly. In Experiments 3 and 4, subjects showed implicit learning when the task involved either motor responses to verbal stimuli or verbal responses to spatially arranged stimuli. These findings are discussed in terms of the conditions under which implicit learning can be obtained. First, they demonstrate implicit learning of a set of new associations in the verbal domain. Second, the data suggest that attention is important in implicit learning. Finally, the degree of interitem organization that is familiar preexperimentally seems to partially determine the amount of implicit learning.     

Incidental learning of real-world regularities

The large literature on incidental learning relies almost exclusively on laboratory experiments. Whenever researchers have attempted to demonstrate incidental learning of real-world regularities, they have typically failed to show learning. For example, it is well established that people do not learn regularities in everyday objects, such as the left-right orientation of faces on coins, despite a very large exposure to them. In this report, we examine this apparent contradiction. We argue that most studies exploring real-life incidental learning use tests that are not as sensitive to low-confidence information as those traditionally used in laboratory tasks. Using more sensitive measures, we show that it is possible to learn regularities from British and Japanese cultural life as a direct result of exposure to these regularities. Further, confidence measures suggest that although the information may be acquired incidentally, it can be expressed with and without concomitant awareness of that knowledge.     

Single-use versus mixed-use learning of transformations: evidence for conceptual integration

The principle of use-specificity proposes that "compiled" knowledge encapsulates the knowledge use in the acquisition context. The efficiency of knowledge use should be a monotone function of use-specific practice. Transfer between different uses of function concepts should decrease with increasing practice. In contrast, the hypothesis of conceptual integration proposes that conceptually integrated knowledge can be used flexibly. The efficiency of knowledge use should be a monotone function of concept-specific practice. The probability to integrate a feature is assumed to depend on its relevance within the acquisition context. Transfer between different uses of function concepts should increase with increasing practice. In Experiment 1, the effects of practicing a single versus two different uses of symbolic functions on learning and subsequent transfer were investigated at 1 day and 2 consecutive days of practice. In Experiment 2, these effects were investigated for 4 days of practice. The findings support the hypothesis of conceptual integration. Performance was mainly a function of concept-specific training. Limited usability of acquired knowledge was rare and disappeared completely at higher levels of practice.