Category learning with minimal prior knowledge

Searching for Category-Consistent Features: A Computational Approach to Understanding Visual Category Representation

This article introduces a generative model of category representation that uses computer vision methods to extract category-consistent features (CCFs) directly from images of category exemplars. The model was trained on 4,800 images of common objects, and CCFs were obtained for 68 categories spanning subordinate, basic, and superordinate levels in a category hierarchy. When participants searched for these same categories, targets cued at the subordinate level were preferentially fixated, but fixated targets were verified faster when they followed a basic-level cue. The subordinate-level advantage in guidance is explained by the number of target-category CCFs, a measure of category specificity that decreases with movement up the category hierarchy. The basic-level advantage in verification is explained by multiplying the number of CCFs by sibling distance, a measure of category distinctiveness. With this model, the visual representations of real-world object categories, each learned from the vast numbers of image exemplars accumulated throughout everyday experience, can finally be studied.

Category vs. Object Knowledge in Category-based Induction

In one form of category-based induction, people make predictions about unknown properties of objects. There is a tension between predictions made based on the object's specific features (e.g., objects above a certain size tend not to fly) and those made by reference to category-level knowledge (e.g., birds fly). Seven experiments with artificial categories investigated these two sources of induction by looking at whether people used information about correlated features within categories, suggesting that they focused on feature-feature relations rather than summary categorical information. The results showed that people relied heavily on such correlations, even when there was no reason to think that the correlations exist in the population. The results suggested that people's use of this strategy is largely unreflective, rather than strategically chosen. These findings have important implications for models of category-based induction, which generally ignore feature-feature relations.

How does prior knowledge affect implicit and explicit concept learning?

Two experiments investigated the effect of prior knowledge on implicit and explicit learning. Implicit as opposed to explicit learning is sometimes characterized as unselective or purely statistical. During training, one group of participants was presented with category exemplars whose features could be tied together by integrative knowledge, whereas another group saw category exemplars with unrelated feature combinations. Half of the participants in each group learned these categories under a secondary-task condition (meant to discourage explicit learning), and the remaining half performed the categorization task under a single-task condition (meant to favour explicit learning). In a test phase, participants classified only the individual features of the training exemplars. Secondary- as opposed to single-task conditions impaired explicit but not implicit knowledge (as determined by subjective measures). Importantly, prior knowledge resulted in increased amounts of both implicit and explicit knowledge.

Prior knowledge enhances the category dimensionality effect

A study of the combined influence of prior knowledge and stimulus dimensionality on category learning was conducted. Subjects learned category structures with the same number of necessary dimensions but with more or fewer additional, redundant dimensions and with either knowledge-related or knowledge-unrelated features. Minimal-learning models predict that all subjects, regardless of condition, either should learn the same number of dimensions or should respond more slowly to each dimension. Despite similar learning rates and response times, subjects learned more features in the high-dimensional than in the low-dimensional condition. Furthermore, prior knowledge interacted with dimensionality, increasing what was learned, especially in the high-dimensional case. A second experiment confirmed that the participants did, in fact, learn more features during the training phase, rather than simply inferring them at test. These effects can be explained by direct associations among features (representing prior knowledge), combined with feedback between features and the category label, as was shown by simulations of the knowledge resonance, or KRES, model of category learning.

Subtyping as a knowledge preservation strategy in category learning

Subtyping occurs when atypical examples are excluded from consideration in judging a category. In three experiments, we investigated whether subtyping can influence category learning. In each experiment, participants learned about a category where most, but not all, of the exemplars corresponded to a theme. The category structure included a subtyping dimension, which had one value for theme-congruent exemplars and another for exception exemplars. On the basis of work by Hayes, Foster, and Gadd (2003) and studies on social stereotyping, we hypothesized that this subtyping dimension would enable the participants to discount the exception exemplars, thereby facilitating category learning. Contrary to expectations, we did not find a subtyping effect with traditional category-learning procedures. By introducing the theme prior to learning, however, we observed increased effects on typicality ratings (Experiment 1) and facilitated learning of the category (Experiment 2). Experiment 3 provided direct evidence that introducing the theme prior to learning enhanced the subtyping effect and provided support for a knowledge-gating explanation of subtyping. We conclude that subtyping effects are strongest on already-learned concepts and that subtyping is unlikely to aid in the learning of new concepts, except in particular circumstances.

Ad hoc category restructuring

Participants learned to classify seemingly arbitrary words into categories that also corresponded to ad hoc categories (see, e.g., Barsalou, 1983). By adapting experimental mechanisms previously used to study knowledge restructuring in perceptual categorization, we provide a novel account of how experimental and preexperimental knowledge interact. Participants were told of the existence of the ad hoc categories either at the beginning or the end of training. When the ad hoc labels were revealed at the end of training, participants switched from categorization based on experimental learning to categorization based on preexperimental knowledge in some, but not all, circumstances. Important mediators of the extent of that switch were the amount of performance error experienced during prior learning and whether or not prior knowledge was in conflict with experimental learning. We present a computational model of the trade-off between preexperimental knowledge and experimental learning that accounts for the main results.

The science of cycology: Failures to understand how everyday objects work

When their understanding of the basics of bicycle design was assessed objectively, people were found to make frequent and serious mistakes, such as believing that the chain went around the front wheel as well as the back wheel. Errors were reduced but not eliminated for bicycle experts, for men more than women, and for people who were shown a real bicycle as they were tested. The results demonstrate that most people's conceptual understanding of this familiar, everyday object is sketchy and shallow, even for information that is frequently encountered and easily perceived. This evidence of a minimal and even inaccurate causal understanding is inconsistent with that of strong versions of explanation-based (or theory-based) theories of categorization.

Instantiated features and the use of "rules."

Classification "rules" in expert and everyday discourse are usually deficient by formal standards, lacking explicit decision procedures and precise terms. The authors argue that a central function of such weak rules is to focus on perceptual learning rather than to provide definitions. In 5 experiments, transfer following learning of family resemblance categories was influenced more by familiar-appearing features than by novel-appearing features equally acceptable under the rule. This occurred both when rules were induced and when rules were given at the beginning of instruction. To model this and other phenomena in categorization, features must be represented on 2 levels: informational and instantiated. These 2 feature levels are crucial to provide broad generalization while reflecting the known peculiarities of a complex world.

Expectations about stimulus structure in implicit learning

In two experiments, artificial grammar learning was used to create learning tasks of identical symbolic complexity that differed in terms of the instructions and the appearance of the stimuli. In Experiment 1, the stimuli were letter strings that appeared in either upper- or lowercase; however, letter case was irrelevant to the learning task. Participants were unable to take into account the instruction to ignore letter case. In Experiment 2, the stimuli were sequences of cities that corresponded to the routes of a traveling salesman. It was assumed that participants would adopt the expectation that the salesman would prefer short journeys. When the structure of the stimuli was inconsistent with this expectation, performance on the learning task was inhibited. The results suggest that there are circumstances in which explicit expectations about the structure of a set of stimuli can affect implicit learning of the stimuli.

Category coherence and category-based property induction

One important property of human object categories is that they define the sets of exemplars to which newly observed properties are generalized. We manipulated the causal knowledge associated with novel categories and assessed the resulting strength of property inductions. We found that the theoretical coherence afforded to a category by inter-feature causal relationships strengthened inductive projections. However, this effect depended on the degree to which the exemplar with the to-be-projected predicate manifested or satisfied its category's causal laws. That is, the coherence that supports inductive generalizations is a property of individual category members rather than categories. Moreover, we found that an exemplar's coherence was mediated by its degree of category membership. These results were obtained across a variety of causal network topologies and kinds of categories, including biological kinds, non-living natural kinds, and artifacts.

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.

Putting together prior knowledge, verbal arguments, and observations in category learning

Two experiments addressed the novel issue of how people incorporate verbal arguments into category learning. In Experiment 1, at the start of learning, subjects were given verbal arguments, which had an influence equivalent to a fixed number of category members. In Experiment 2, subjects learned under slower paced conditions, and it was found that both prior knowledge and arguments had multiple effects on categorization: a fixed initial influence plus selective weighting of new observations. The results supported the idea that verbally presented arguments can be treated in a similar manner as other forms of prior knowledge, from the perspective of applying models of categorization.