Discrimination of structure: I. Implications for connectionist theories of discrimination learning

In each of 4 experiments animals were given a structural discrimination task that involved visual patterns composed of identical features, but the spatial relations among the features were different for reinforced and nonreinforced trials. In Experiment 1 the stimuli were pairs of colored circles, and pigeons were required to discriminate between patterns that were the mirror image of each other. A related task was given to rats in Experiment 2. Subjects solved these discriminations. For Experiment 3, some pigeons were given a discrimination similar to that used in Experiment 1, which they solved, whereas others received a comparable task but with 3 colored circles present on every trial, which they failed to solve. The findings from Experiment 3 were replicated in Experiment 4 using different patterns. The results are difficult to explain by certain connectionist theories of discrimination learning, unless they are modified to take account of the way in which compound stimuli are structured.     

ALCOVE: an exemplar-based connectionist model of category learning.

ALCOVE (attention learning covering map) is a connectionist model of category learning that incorporates an exemplar-based representation (Medin & Schaffer, 1978; Nosofsky, 1986) with error-driven learning (Gluck & Bower, 1988; Rumelhart, Hinton, & Williams, 1986). Alcove selectively attends to relevant stimulus dimensions, is sensitive to correlated dimensions, can account for a form of base-rate neglect, does not suffer catastrophic forgetting, and can exhibit 3-stage (U-shaped) learning of high-frequency exceptions to rules, whereas such effects are not easily accounted for by models using other combinations of representation and learning method.     

Effects of similarity on apparent motion and perceptual grouping

Effects of similarity in colour, luminance, size, and shape on apparent motion and perceptual grouping were examined in part 1 in two parallel experiments on the same seven subjects. In both experiments, the effect of similarity was compared with that of proximity in competitive, bistable stimulus situations. A combination of a larger horizontal separation between the homogeneous stimulus elements and a smaller constant vertical separation between heterogeneous stimulus elements produced two kinds of apparent motion (or perceptual grouping) with equal probabilities. Such matched separations between homogeneous stimulus elements were obtained by the double staircase method in various stimulus conditions. In both experiments on apparent motion and perceptual grouping matched separation was found to increase as the difference between the heterogeneous stimulus elements increased. High correlations (0.71 to 0.94) of matched separations were found between apparent motion and perceptual grouping in four stimulus series: colour, luminance, size, and shape. Six of the seven subjects were also tested in part 2. Here, the effects of differences were found to work additively across different perceptual attributes in both phenomena, when multiple differences were combined in heterogeneous elements. The experimental results are discussed from the point of view that apparent motion is an example of perceptual constancy.     

Effects of preexposure on stimulus discrimination: An investigation of the mechanisms responsible for human perceptual learning

The effect of preexposure on human perceptual learning was investigated in four experiments. In Experiments 1a and 1b, participants were preexposed to one pair of visual stimuli on an intermixed schedule (AX/BX) and one on a blocked schedule (CX_DX). The ability to discriminate between AX and BX and between CX and DX was then assessed by examining the extent to which key presses assigned to each member of a stimulus pair generalized to the other member (Experiment 1a) and by looking at the accuracy of same–different responses (Experiment 1b). Stimuli were more easily discriminated following intermixed than following blocked preexposure on both the generalization and same–different tasks. This suggests that two stimuli are more perceptually distinct after intermixed preexposure. Experiments 2a and 2b investigated the mechanisms responsible for perceptual learning using same–different tasks. The results support the suggestion that the enhanced discrimination observed after intermixed preexposure is due to increases in the salience of the unique elements.     

Effects of preexposure on stimulus discrimination: an investigation of the mechanisms responsible for human perceptual learning

The effect of preexposure on human perceptual learning was investigated in four experiments. In Experiments 1a and 1b, participants were preexposed to one pair of visual stimuli on an intermixed schedule (AX/BX) and one on a blocked schedule (CX_DX). The ability to discriminate between AX and BX and between CX and DX was then assessed by examining the extent to which key presses assigned to each member of a stimulus pair generalized to the other member (Experiment 1a) and by looking at the accuracy of same-different responses (Experiment 1b). Stimuli were more easily discriminated following intermixed than following blocked preexposure on both the generalization and same-different tasks. This suggests that two stimuli are more perceptually distinct after intermixed preexposure. Experiments 2a and 2b investigated the mechanisms responsible for perceptual learning using same-different tasks. The results support the suggestion that the enhanced discrimination observed after intermixed preexposure is due to increases in the salience of the unique elements.     

Effects of stress on nonassociative learning processes in male and female rats

In this study we assessed habituation and sensitization of the acoustic startle response (ASR) to discern whether intense, inescapable stress affects nonassociative learning differently in male and female rats. Rats were inescapably stressed 2 hours per day over 3 consecutive days. ASR magnitudes were measured at several times post-stress (1, 4, 8, and 15 days after cessation). Females generally showed greater ASR magnitudes (compared to males), but both sexes exhibited short and long-term habituation across the testing days. ASR magnitudes were only affected by stress in male subjects. The effect in males was an increase in short-term sensitization of the ASR on post-stress day-4. The results suggest that stressed males and females react differently to ASR testing, in that stress males appear to develop an exaggerated ASR response over repeated test sessions due to short-term sensitization. The source of the short-term sensitization is discussed with regards to possible stress-induced enhanced contextual learning during ASR testing on post-stress day-1.     

Effects of perceptual and conceptual similarity in semantic priming

Summary The study investigates the different contributions to semantic priming of two components of the semantic representation underlying a word. The two components are perceptually based information and conceptually based information. Perceptual information is based on physical attributes such as shape or color, while conceptual information consists of more abstract elements such as functional attributes. The question asked in this study was whether both components would produce an effect in semantic priming. Pairs of words either related because of a conceptual property (banana-apple), a perceptual property (ball-apple), or both because of a perceptual and conceptual property (cherry-apple) were presented as prime and target in a lexical decision and a word-naming task. The results showed independent contributions of perceptual and conceptual attributes to semantic priming.This research was supported by a grant from the Dutch Organization for the Advancement of Pure Research (ZWO). All three authors are senior authors and have contributed to all parts of the project.     

Temporal specificity of perceptual learning in an auditory discrimination task

Although temporal processing is used in a wide range of sensory and motor tasks, there is little evidence as to whether a single centralized clock or a distributed system underlies timing in the range of tens to hundreds of milliseconds. We investigated this question by studying whether learning on an auditory interval discrimination task generalizes across stimulus types, intervals, and frequencies. The degree to which improvements in timing carry over to different stimulus features constrains the neural mechanisms underlying timing. Human subjects trained on a 100- or 200-msec interval discrimination task showed an improvement in temporal resolution. This learning generalized to a perceptually distinct duration stimulus, as well as to the trained interval presented with tones at untrained spectral frequencies. The improvement in performance did not generalize to untrained intervals. To determine if spectral generalization was dependent on the importance of frequency information in the task, subjects were simultaneously trained on two different intervals identified by frequency. As a whole, our results indicate that the brain uses circuits that are dedicated to specific time spans, and that each circuit processes stimuli across nontemporal stimulus features. The patterns of generalization additionally indicate that temporal learning does not rely on changes in early, subcortical processing, because the nontemporal features are encoded by different channels at early stages.     

Recalibration by pairing: a process of perceptual learning.

When two cues which normally agree in determining perceived depth are experimentally paired so that they designate discrepant values for the depth dimension, an opportunity is created for the recalibration or re-evaluation of the cues. Experiments which show recalibration of the oculomotor cues and recalibration of binocular disparity as a result of pairing these cues with other discrepant cues are reviewed. The locus of change is related to the relative dominance of different cues and the contribution of allocation of attention in determining dominance is discussed. Implications of recalibration by pairing for understanding ontogenetic development are considered. It is suggested that some cues may acquire their effectiveness as a consequence of pairing during development.     

A feedback model of perceptual learning and categorization

Top-down, feedback, influences are known to have significant effects on visual information processing. Such influences are also likely to affect perceptual learning. This paper employs a computational model of the cortical region inter-actions underlying visual perception to investigate possible influences of top-down information on learning. The results suggest that feedback could bias the way in which perceptual stimuli are categorized and could also facilitate the learning of subordinate level representations suitable for object identification and perceptual expertise.     

Combining exemplar-based category representations and connectionist learning rules.

Adaptive network and exemplar-similarity models were compared on their ability to predict category learning and transfer data. An exemplar-based network (Kruschke, 1990a, 1990b, 1992) that combines key aspects of both modeling approaches was also tested. The exemplar-based network incorporates an exemplar-based category representation in which exemplars become associated to categories through the same error-driven, interactive learning rules that are assumed in standard adaptive networks. Experiment 1, which partially replicated and extended the probabilistic classification learning paradigm of Gluck and Bower (1988a), demonstrated the importance of an error-driven learning rule. Experiment 2, which extended the classification learning paradigm of Medin and Schaffer (1978) that discriminated between exemplar and prototype models, demonstrated the importance of an exemplar-based category representation. Only the exemplar-based network accounted for all the major qualitative phenomena; it also achieved good quantitative predictions of the learning and transfer data in both experiments.     

Lesioning a connectionist model of visual search: selective effects on distractor grouping.

A connectionist model of visual search is presented, in which search is determined by patterns of grouping between distractors and between the target and the distractors. Grouping is based on conjunctive relations between simple form elements (corners, line-end terminators), and uses principles of similarity and spatial proximity. In its normal, 'unlesioned' state, the model simulates the search data generated by human subjects when they search for simple form conjunctions amongst either homogeneous or heterogeneous distractors (Humphreys & Muller, in press). In this paper, the performance of the model is examined after the model is subject to various types of 'lesion'. 'Lesioning' is produced either by increasing the internal noise on the activation functions governing the interactions between processing units, or by eliminating processing units from different loci in the model. Increasing the internal noise within the model, or lesioning units within 'high-level' processing stages, can generate the selective effects on search found in a detailed single case study of an agnosic patient (Humphreys, Riddoch, Quinlan, Price & Donnelly, this volume): Namely, there is selective disruption of search for a form conjunction amongst homogeneous distractors relative to search for the same target amongst heterogeneous distractors. This selective effect can be attributed to the 'lesion' disrupting grouping between distractors, which normally facilitates search with homogeneous distractors. The simulations demonstrate that (1) search processes can be disrupted by adding internal noise to search functions, (2) that there can be selective effects on grouping processes, and (3) that these selective effects can be associated with different types of lesion to different stages in the model. The implications of the simulations for understanding visual processing impairments in agnosia are discussed.     

Symbolically speaking: a connectionist model of sentence production

The ability to combine words into novel sentences has been used to argue that humans have symbolic language production abilities. Critiques of connectionist models of language often center on the inability of these models to generalize symbolically (Fodor & Pylyshyn, 1988; Marcus, 1998). To address these issues, a connectionist model of sentence production was developed. The model had variables (role‐concept bindings) that were inspired by spatial representations (Landau & Jackendoff, 1993). In order to take advantage of these variables, a novel dual‐pathway architecture with event semantics is proposed and shown to be better at symbolic generalization than several variants. This architecture has one pathway for mapping message content to words and a separate pathway that enforces sequencing constraints. Analysis of the model's hidden units demonstrated that the model learned different types of information in each pathway, and that the model's compositional behavior arose from the combination of these two pathways. The model's ability to balance symbolic and statistical behavior in syntax acquisition and to model aphasic double dissociations provided independent support for the dual‐pathway architecture.     

The dynamics of perceptual learning: an incremental reweighting model

The mechanisms of perceptual learning are analyzed theoretically, probed in an orientation-discrimination experiment involving a novel nonstationary context manipulation, and instantiated in a detailed computational model. Two hypotheses are examined: modification of early cortical representations versus task-specific selective reweighting. Representation modification seems neither functionally necessary nor implied by the available psychophysical and physiological evidence. Computer simulations and mathematical analyses demonstrate the functional and empirical adequacy of selective reweighting as a perceptual learning mechanism. The stimulus images are processed by standard orientation- and frequency-tuned representational units, divisively normalized. Learning occurs only in the "read-out" connections to a decision unit; the stimulus representations never change. An incremental Hebbian rule tracks the task-dependent predictive value of each unit, thereby improving the signal-to-noise ratio of their weighted combination. Each abrupt change in the environmental statistics induces a switch cost in the learning curves as the system temporarily works with suboptimal weights.