Online research in philosophy

Schyns, Goldstone & Thibaut present reasonable arguments for feature creation in category learning. We argue, however, that they do not provide unequivocal evidence either for the necessity or for the occurrence of feature creation. In an effort to sharpen the debate, we take the stand that a fixed feature approach is to be preferred in the absence of compelling evidence.

If we think of perceptual expertise, we might think ofa neurologist interpreting a CAT scan or an astronomerlooking at a star. But perceptual expertise is notlimited to experts. Perceptual expertise is atthe heart of our everyday competence in the world. Wenavigate around obstacles, we take turns inconversations, we make left-turns in face of on-comingtraffic. Each of us is a perceptual expert (thoughonly in certain domains). If we misunderstandperceptual expertise, we risk misunderstanding ourepistemic relationship to the world. I argue that thestandard arguments for the received view of perceptualexpertise are problematic at best. Of course, theissue of whether the received view is actually correctis an empirical issue. But the decision to adopt thereceived view, I argue, was not a scientific decision,but was made by inheriting a philosophical tradition– which many philosophers today would question.

Cognitive theories based on a fixed feature set suffer from frame and symbol grounding problems. Flexible features and other empirically acquired constraints (e.g., analog-to-analog mappings) provide a framework for letting extrinsic relations influence symbol manipulation. By offering a biologically plausible basis for feature learning, nonorthogonal multiresolution analysis and dimensionality reduction, informed by functional constraints, may contribute to a solution to the symbol grounding problem.

We report the results of an experiment in which human subjects were trained to perform a perceptual matching task. Subjects were asked to manipulate comparison objects until they matched target objects using the fewest manipulations possible. An unusual feature of the experimental task is that efficient performance requires an understanding of the hidden or latent causal structure governing the relationships between actions and perceptual outcomes. We use two benchmarks to evaluate the quality of subjects’ learning. One benchmark is based on optimal performance as calculated by a dynamic programming procedure. The other is based on an adaptive computational agent that uses a reinforcement-learning method known as Q-learning to learn to perform the task. Our analyses suggest that subjects were successful learners. In particular, they learned to perform the perceptual matching task in a near-optimal manner (i.e., using a small number of manipulations) at the end of training. Subjects were able to achieve near-optimal performance because they learned, at least partially, the causal structure underlying the task. In addition, subjects’ performances were broadly consistent with those of model-based reinforcement-learning agents that built and used internal models of how their actions influenced the external environment. We hypothesize that people will achieve near-optimal performances on tasks requiring sequences of action—especially sensorimotor tasks with underlying latent causal structures—when they can detect the effects of their actions on the environment, and when they can represent and reason about these effects using an internal mental model.

The origin of features from nonfeatural information is a problem that should concern all theories of object categorization and recognition, not just the flexible feature approach. In contrast to the idea that new features must originate from combinations of simpler fixed features, we argue that holistic features can be created from a direct imprinting on the visual medium. Furthermore, featural descriptions can emerge from processes that by themselves do not operate on feature detectors. Once acquired, features can be decomposed into component features if required by other categorizations. We therefore argue that it is not necessary to separate holistic and componential approaches to representations, because the latter is a development of the former. The requirements for representational flexibility outstrip the performance of any existing computational models, but specific mechanisms of feature creation are discussed and evaluated. Challenges for feature creation mechanisms are discussed together with the constraints (perceptual, statistical, functional, and task) they will need to satisfy.

We introduce an innovative technique that quantifies human expertise development in such a way that humans and artificial systems can be directly compared. Using this technique we are able to highlight certain fundamental difficulties associated with the learning of a complex task that humans are still exceptionally better at than their computer counterparts. We demonstrate that expertise goes through significant developmental transitions that have previously been predicted but never explicated. The first signals the onset of a steady increase in global awareness that begins surprisingly late in expertise acquisition. The second transition, reached by only a very few experts in the world, shows a major reorganisation of global contextual knowledge resulting in a relatively minor gain in skill. We are able to show that these empirical findings have consequences for our understanding of the way in which expertise acquisition may be modelled by learning in artificial intelligence systems. This point is emphasised with a novel theoretical result showing explicitly how our findings imply a non-trivial hurdle for learning for suitably complex tasks.

Abstract Since their inception, distributional models of semantics have been criticized as inadequate cognitive theories of human semantic learning and representation. A principal challenge is that the representations derived by distributional models are purely symbolic and are not grounded in perception and action; this challenge has led many to favor feature-based models of semantic representation. We argue that the amount of perceptual and other semantic information that can be learned from purely distributional statistics has been underappreciated. We compare the representations of three feature-based and nine distributional models using a semantic clustering task. Several distributional models demonstrated semantic clustering comparable with clustering-based on feature-based representations. Furthermore, when trained on child-directed speech, the same distributional models perform as well as sensorimotor-based feature representations of children’s lexical semantic knowledge. These results suggest that, to a large extent, information relevant for extracting semantic categories is redundantly coded in perceptual and linguistic experience. Detailed analyses of the semantic clusters of the feature-based and distributional models also reveal that the models make use of complementary cues to semantic organization from the two data streams. Rather than conceptualizing feature-based and distributional models as competing theories, we argue that future focus should be on understanding the cognitive mechanisms humans use to integrate the two sources.

Schyns, Goldstone & Thibaut argue that categorization experience results in the learning of new perceptual features that are not derivable from the learner's existing feature set. We explore the meaning and implications of this “nonderivability” claim and relate it to the question of whether perceptual invariants are learnable, and if so, what might be entailed in learning them.

According to one productive and influential approach to cognition, categorization, object recognition, and higher level cognitive processes operate on a set of fixed features, which are the output of lower level perceptual processes. In many situations, however, it is the higher level cognitive process being executed that influences the lower level features that are created. Rather than viewing the repertoire of features as being fixed by low-level processes, we present a theory in which people create features to subserve the representation and categorization of objects. Two types of category learning should be distinguished. Fixed space category learning occurs when new categorizations are representable with the available feature set. Flexible space category learning occurs when new categorizations cannot be represented with the features available. Whether fixed or flexible, learning depends on the featural contrasts and similarities between the new category to be represented and the individual's existing concepts. Fixed feature approaches face one of two problems with tasks that call for new features: If the fixed features are fairly high level and directly useful for categorization, then they will not be flexible enough to represent all objects that might be relevant for a new task. If the fixed features are small, subsymbolic fragments (such as pixels), then regularities at the level of the functional features required to accomplish categorizations will not be captured by these primitives. We present evidence of flexible perceptual changes arising from category learning and theoretical arguments for the importance of this flexibility. We describe conditions that promote feature creation and argue against interpreting them in terms of fixed features. Finally, we discuss the implications of functional features for object categorization, conceptual development, chunking, constructive induction, and formal models of dimensionality reduction. Key Words: concept learning; conceptual development; features; perceptual learning; stimulus encoding.

The fixed-feature viewpoint Schyns et al. are opposing is not a widely held theoretical position but rather a working assumption of cognitive psychologists – and thus a straw man. We accept their demonstration of new-feature acquisition, but question its ubiquity in category learning. We suggest that new-feature learning (at least in adults) is rarer and more difficult than the authors suggest.