Mechanistic explanations satisfy widely held norms of explanation: the ability to manipulate and answer counterfactual questions about the explanandum phenomenon. A currently debated issue is whether any nonmechanistic explanations can satisfy these explanatory norms. Weiskopf argues that the models of object recognition and categorization, JIM, SUSTAIN, and ALCOVE, are not mechanistic yet satisfy these norms of explanation. In this article I argue that these models are mechanism sketches. My argument applies recent research using model-based functional magnetic resonance imaging, a novel (...) neuroimaging method whose significance for current debates on psychological models and mechanistic explanation has yet to be explored. (shrink)

The distinction between rules and similarity is central to our understanding of much of cognitive psychology. Two aspects of existing research have motivated the present work. First, in different cognitive psychology areas we typically see different conceptions of rules and similarity; for example, rules in language appear to be of a different kind compared to rules in categorization. Second, rules processes are typically modeled as separate from similarity ones; for example, in a learning experiment, rules and similarity influences would be (...) described on the basis of separate models. In the present article, I assume that the rules versus similarity distinction can be understood in the same way in learning, reasoning, categorization, and language, and that a unified model for rules and similarity is appropriate. A rules process is considered to be a similarity one where only a single or a small subset of an object's properties are involved. Hence, rules and overall similarity operations are extremes in a single continuum of similarity operations. It is argued that this viewpoint allows adequate coverage of theory and empirical findings in learning, reasoning, categorization, and language, and also a reassessment of the objectives in research on rules versus similarity. Key Words: categorization; cognitive explanation; language; learning; reasoning; rules; similarity. (shrink)

We argue that atomistic learning?learning that requires training only on a novel item to be learned?is problematic for networks in which every weight is available for change in every learning situation. This is potentially significant because atomistic learning appears to be commonplace in humans and most non-human animals. We briefly review various proposed fixes, concluding that the most promising strategy to date involves training on pseudo-patterns along with novel items, a form of learning that is not strictly atomistic, but which (...) looks very much like it ?from the outside? (shrink)

Page's proposal to stipulate representations in which individual units correspond to meaningful entities is too unconstrained to support effective theorizing. An approach combining general computational principles with domain-specific assumptions, in which learning is used to discover representations that are effective in solving tasks, provides more insight into why cognitive and neural systems are organized the way they are.

Amit's “Attractor Neural Network” perspective on cognition raises difficult technical problems already met by prior dynamical models. This commentary sketches briefly some of them concerning the internal topological structure of attractors, the constituency problem, the possibility of activating simultaneously several attractors, and the different kinds of dynamical structures one can use to model brain activity: point attractors, strange attractors, synchronized arrays of oscillators, synfire chains, and so forth.

Decades of psychological research have been aimed at modeling how people learn features and categories. The empirical validation of these theories is often based on artificial stimuli with simple representations. Recently, deep neural networks have reached or surpassed human accuracy on tasks such as identifying objects in natural images. These networks learn representations of real‐world stimuli that can potentially be leveraged to capture psychological representations. We find that state‐of‐the‐art object classification networks provide surprisingly accurate predictions of human similarity judgments for (...) natural images, but they fail to capture some of the structure represented by people. We show that a simple transformation that corrects these discrepancies can be obtained through convex optimization. We use the resulting representations to predict the difficulty of learning novel categories of natural images. Our results extend the scope of psychological experiments and computational modeling by enabling tractable use of large natural stimulus sets. (shrink)

Inductive generalization, where people go beyond the data provided, is a basic cognitive capability, and it underpins theoretical accounts of learning, categorization, and decision making. To complete the inductive leap needed for generalization, people must make a key ‘‘sampling’’ assumption about how the available data were generated. Previous models have considered two extreme possibilities, known as strong and weak sampling. In strong sampling, data are assumed to have been deliberately generated as positive examples of a concept, whereas in weak sampling, (...) data are assumed to have been generated without any restrictions. We develop a more general account of sampling that allows for an intermediate mixture of these two extremes, and we test its usefulness. In two experiments, we show that most people complete simple one-dimensional generalization tasks in a way that is consistent with their believing in some mixture of strong and weak sampling, but that there are large individual differences in the relative emphasis different people give to each type of sampling. We also show experimentally that the relative emphasis of the mixture is influenced by the structure of the available information. We discuss the psychological meaning of mixing strong and weak sampling, and possible extensions of our modeling approach to richer problems of inductive generalization. (shrink)

Linguistic and non-linguistic pattern learning have been studied separately, but we argue for a comparative approach. Analogous inductive problems arise in phonological and visual pattern learning. Evidence from three experiments shows that human learners can solve them in analogous ways, and that human performance in both cases can be captured by the same models. We test GMECCS, an implementation of the Configural Cue Model in a Maximum Entropy phonotactic-learning framework with a single free parameter, against the alternative hypothesis that learners (...) seek featurally simple algebraic rules. We study the full typology of patterns introduced by Shepard, Hovland, and Jenkins (), instantiated as both phonotactic patterns and visual analogs, using unsupervised training. Unlike SHJ, Experiments 1 and 2 found that both phonotactic and visual patterns that depended on fewer features could be more difficult than those that depended on more features, as predicted by GMECCS but not by rule-seeking. GMECCS also correctly predicted performance differences between stimulus subclasses within each pattern. A third experiment tried supervised training to elicit simple rule-seeking phonotactic learning, but cue-based behavior persisted. We conclude that similar cue-based cognitive processes are available for phonological and visual concept learning, and hence that studying either kind of learning can lead to significant insights about the other. (shrink)

The Miyashita experiments are very interesting and the results should be examined from a viewpoint of attractor dynamics. Amit's target article shows a path toward realistic modeling by artificial neural networks (ANN), but it is not necessarily the only one. I introduce another model that can explain a substantial part of the empirical observations and makes an interesting prediction. This model consists of such units that have nonmonotonic input-output characteristics with local inhibition neurons.

This article presents a process account of some typicality effects and related similarity-dependent accuracy and response time phenomena that arise in the context of supervised concept acquisition. We describe Symbolic Concept Acquisition (SCA), a computational system that acquires and activates category prediction rules. In contrast to gradient representations, SCA performs by probing for prediction rules in a series of discrete steps. For learning new rules, it acquires general rules but then incrementally learns more specific ones. In describing SCA, we emphasize (...) its functionality in terms of accuracy and efficiency and motivate its design within the set of symbolic mechanisms and memory structures defined by the Soar architecture (Laird, Newell & Rosenbloom, 1987). For replicating human behavior, we first show how SCA exhibits some typicality effects in the course of learning responding faster and more accurately to more typical test examples. Then, using data from human experiments, we evaluate SCA's qualitative predictions on accuracy and response time on individual dataset instances. We show how SCA's predictions correlate with human data across three experimental conditions concerning the effect of instruction on learning strategy. (shrink)

Every immediate memory is unique; it is therefore unlikely to consist of an attractor or even a combination of attractors. In the present state of knowledge about the chemistry of synaptic transmission, there is no reason to look beyond neurons that directly receive sensory afferents for the afterdischarges that correspond to active memories.

This article addresses the ability of Parallel Distributed Processing (PDP) networks to generate stagewise cognitive development in accordance with Piaget's theory of cognitive epigenesis. We carried out a replication study of the simulation experiments by McClelland (1989) and McClelland and Jenkins (1991) in which a PDP network learns to solve balance scale problems. In objective tests motivated from catastrophe theory, a mathematical theory of transitions in epigenetical systems, no evidence for stage transitions in network performance was found. It is concluded (...) that PDP networks lack the ability to recover the positive outcomes of analogous catastrophe analyses of real cognitive developmental data. In an attempt to further characterize the learning behaviour of PDP networks, we carried out a second simulation study using the discrimination‐shift paradigm. The results thus obtained indicate that PDP learning is compatible with the learning of stimulus‐response relationships, not with the acquisition of mediating rules such as conceived in (neo‐)Piagetian theory. In closing, we speculate about the feasibility of simulating stagewise development with alternative network architectures. (shrink)

This paper presents a skill learning model CLARION. Different from existing models of mostly high-level skill learning that use a top-down approach (that is, turning declarative knowledge into procedural knowledge through practice), we adopt a bottom-up approach toward low-level skill learning, where procedural knowledge develops first and declarative knowledge develops later. Our model is formed by integrating connectionist, reinforcement, and symbolic learning methods to perform on-line reactive learning. It adopts a two-level dual-representation framework (Sun, 1995), with a combination of localist (...) and distributed representation. We compare the model with human data in a minefield navigation task, demonstrating some match between the model and human data in several respects. (shrink)

Rogers & McClelland's (R&M's) précis represents an important effort to address key issues in concepts and categorization, but few of the simulations deliver what is promised. We argue that the models are seriously underconstrained, importantly incomplete, and psychologically implausible; more broadly, R&M dwell too heavily on the apparent successes without comparable concern for limitations already noted in the literature.

Algorithms for approximate Bayesian inference, such as those based on sampling (i.e., Monte Carlo methods), provide a natural source of models of how people may deal with uncertainty with limited cognitive resources. Here, we consider the idea that individual differences in working memory capacity (WMC) may be usefully modeled in terms of the number of samples, or “particles,” available to perform inference. To test this idea, we focus on two recent experiments that report positive associations between WMC and two distinct (...) aspects of categorization performance: the ability to learn novel categories, and the ability to switch between different categorization strategies (“knowledge restructuring”). In favor of the idea of modeling WMC as a number of particles, we show that a single model can reproduce both experimental results by varying the number of particles—increasing the number of particles leads to both faster category learning and improved strategy‐switching. Furthermore, when we fit the model to individual participants, we found a positive association between WMC and best‐fit number of particles for strategy switching. However, no association between WMC and best‐fit number of particles was found for category learning. These results are discussed in the context of the general challenge of disentangling the contributions of different potential sources of behavioral variability. (shrink)

This article demonstrates the potential of using hierarchical Bayesian methods to relate models and data in the cognitive sciences. This is done using a worked example that considers an existing model of category representation, the Varying Abstraction Model (VAM), which attempts to infer the representations people use from their behavior in category learning tasks. The VAM allows for a wide variety of category representations to be inferred, but this article shows how a hierarchical Bayesian analysis can provide a unifying explanation (...) of the representational possibilities using 2 parameters. One parameter controls the emphasis on abstraction in category representations, and the other controls the emphasis on similarity. Using 30 previously published data sets, this work shows how inferences about these parameters, and about the category representations they generate, can be used to evaluate data in terms of the ongoing exemplar versus prototype and similarity versus rules debates in the literature. Using this concrete example, this article emphasizes the advantages of hierarchical Bayesian models in converting model selection problems to parameter estimation problems, and providing one way of specifying theoretically based priors for competing models. (shrink)

Hebbian cell-assembly theory and attractor networks are good starting points for modeling cortical processing. Detailed cell models can be useful in understanding the dynamics of attractor networks. Cell assemblies are likely to be distributed, with the cortical column as the local processing unit. Synaptic memory may be dominant in all but the first couple of seconds.

A fundamental question in the study of human cognition is how people learn to predict the category membership of an example from its properties. Leading approaches account for a wide range of data in terms of comparison to stored examples, abstractions capturing statistical regularities, or logical rules. Across three experiments, participants learned a category structure in a low‐dimension, continuous‐valued space consisting of regularly alternating regions of class membership (A B A B). The dependent measure was generalization performance for novel items (...) outside the range of the training space. Human learners often extended the alternation pattern––a finding of critical interest given that leading theories of categorization based on similarity or dimensional rules fail to predict this behavior. In addition, we provide novel theoretical interpretations of the observed phenomenon. (shrink)

Hebbian mechanisms are justified according to their functional utility in an evolutionary sense. The selective advantage of correlating content-contingent stimuli reflects the putative common cause of temporally or spatially contiguous inputs. The selective consequences of such correlations are discussed by using examples from the evolution of signal form in sexual selection and model-mimic coevolution. We suggest that evolutionary justification might be considered in addition to neurophysiology plansibility when constructing representational models.

Amit argues that the local neuronal spike rate that persists (reverberating) in the absence of the eliciting stimulus represents the code of the eliciting stimulus. Based on the general argument that the inferred functional meaning of reverberation depends in part on the type of representational assumptions, reverberations may only be important for the encoding of contextual information.