• 09/07   William Timberlake, Indiana University - abstract
• 09/14   John Kruschke, Indiana University - abstract
• 09/21   Mario Fific, Indiana University - abstract
• 09/28   Ji Son & Linda B. Smith, Indiana University - abstract
• 10/05   Richard Shiffrin, Indiana University - abstract
• 10/12   Hans Colonius, University of Oldenburg - abstract
• 10/19   Mark Mon-Williams, University of Aberdeen - abstract
• 10/26   Jennifer Zapf, Indiana University - abstract
• 11/02   Jeroen Raaijmakers, University of Amsterdam - abstract
• 11/09   No talk (Annual Meeting of the Psychonomic Society)
• 11/16   Todd Gureckis, Indiana University - abstract
• 11/23   No talk (Thanksgiving Break)
• 11/30   Bethany Schneider, Indiana University - abstract
• 12/07   Ryan Jessup, Indiana University - abstract
• 12/14   No talks (Finals)

Abstracts

09/07: William Timberlake
Cognitive Mapping versus Associative Learning in Spatial Tasks:
Research on Cue Interaction (Mapping vs. Blocking)in the Water Maze.

Researchers have been divided for more than half a century over attributing spatial learning in the laboratory to cognitive mapping vs associative mechanisms. Against the background of increasing evidence of multiple mechanisms in niche-related spatial learning, including hypothesized spatial processing "modules," we have been testing the extent to which different stimulus conditions determine one or the other result in a cue interaction paradigm in the water maze.

09/14: John Kruschke
Locally Bayesian Learning with Applications to Retrospective Revaluation and Highlighting

A scheme is described for doing locally Bayesian parameter updating in models structured as directed acyclic graphs of component functions. The basic idea is to back-propagate the target data to interior modules, such that the interior targets are those that maximize the probability of the exterior target. The resulting parameter updating is not globally Bayesian, but can better capture human behavior. The approach is implemented for an associative learning model that first maps inputs to attentionally filtered inputs, and then maps attentionally filtered inputs to outputs. The model is applied to several phenomena exhibited in human learning that have heretofore proven difficult for Bayesian learning models or for associative learning models. The model combines some of the crucial abilities of both approaches. The Bayesian updating allows the associative model to exhibit retrospective revaluation effects such as backward blocking and unovershadowing. The back-propagation of target values to attention allows the model to show trial-order effects, including highlighting and asymmetries between forward and backward blocking.

Paper available at: http://www.indiana.edu/~kruschke/articles/Kruschke_LocalBayes_04July2005.pdf

Holistic face recognition refers to the ability of human cognitive systems to deal in an integrative manner with separate face features. A holistic mental representation of a face is not a simple sum of face parts. It possesses unitary properties and corresponds to the whole face appearance better than to any of its constituent parts. A single face feature is better recognized in the learned face context (e.g. Bill's nose in Bill's face) than in isolation or in a new face context (e.g. Bill's nose in Joe's face; Tanaka & Sengco, 1997). The major goal of this study is to provide a rigorous test of the structure and organization of cognitive processes in the holistic perception of faces. Participants performed in two types of face categorization tasks that utilized either a self-terminating or an exhaustive rule for search (OR and AND conditions). Category membership was determined by the manipulation of two configural properties: eye-separation and lips-position. In the first part of each study, participants learned two groups of faces, and we monitored the changes in the face recognition system architecture and capacity. In the second part, the participants' task was to recognize the learned configurations of face features, presented in different face contexts: in the old learned faces, in a new face background and in isolation. Using the systems factorial theory tests, combined with statistical analyses and model simulations, we were able to reveal the exact organization of the mental processes underlying face perception. The findings supported a view that holism is an emergent property which develops with learning. Overall, processing exhibited a parallel architecture with positive interdependency between features in both the OR and AND conditions. We also found that face units are better recognized in the learned face condition than in both the new face context and isolation conditions. We showed that faces are recognized not as a set of independent face features, but as whole units. We revealed that the cognitive mechanism of positive dependence between face features is responsible for forming holistic faces, and provided a simulation that produced behaviors similar to the experimental observations.

09/28: Ji Son & Linda B. Smith
The work of abstractions for generalization: The case of shape perception development

Object recognition seems to be a product of category learning experiences accrued through development (Smith, 2003). In any learning situation, one must retain and disregard parts of the experience. In learning to recognize objects, young children must parse out relevant shape information and exclude other object features. Studies from development (DeLoache, 1995), cognition (Goldstone & Sakamoto, 2003; Bassok & Holyoak, 1993), and education (Uttal, Liu, & DeLoache, 1995) suggest that novices in particular are easily misled by irrelevant information. Yet a presentation of the relevant principles, abstracted away from less informative content, can effectively produce expert-like responses from novices (Biederman & Shiffrar, 1987; Goldstone & Sakamoto, 2003). We have attempted to apply this to 2-year-olds who are yet novices to object recognition. In two experiments, we have attempted to streamline shape learning by simplifying the learning situation. Children were taught object categories with either simple objects varying only on relevant shape information or complex objects with multiple sources of variation. We found that children were able to generalize according to shape better with simple objects rather than complex ones. Our results suggest that the developmental time course of category generalization depends on the complexity of the learning instances.

Our goal is to identify features used to identify or categorize classes of stimuli (e.g. happy vs. sad faces; good vs. bad stories; fMRI patterns for words vs. nonwords; Gabor patches oriented left or right). Assume data consisting of noisy stimuli partitioned (by people) into two classes; each stimulus consists of values on a large number of dimensions (e.g. pixels making up a visual stimulus; words making up a story; voxels making up an fMRI response). Define a feature as a subset of these dimensions, with specified values. Assume that the human's classification is based on matching each of a set of (unknown, to-be-identified) features to a test stimulus, and combining the results. Our goal is to induce the number of features and their values, given only a large set of specified stimuli and their partitioning into two classes.

Many induction methods are possible, almost all computational, and arising from machine learning. We illustrate with one, Multiple Cause Vector Quantization (MCVQ) developed by David Ross (and Richard Zemel). We consider tasks requiring binary classifications of test stimuli presented in specified external noise. The method is applied first to data from a very simple task run by Jason Gold. Then a simulation study is carried out with a more complex task with multiple and complex features, to illustrate the feasibility of recovering complex features.

We end by asking whether similar methods have been or can be used in analyzing fMRI data.

Multisensory neurons in the deep layers of superior colliculus (DSC) show response enhancement to cross-modal stimuli that are spatially coincident, whereas spatially disparate cross-modal cues either fail to produce multisensory interaction or lead to response depression. However, for unimodal inputs - or when certain cortical influences are eliminated - the response of multisensory DSC neurons parallels that of modality-specific neurons. Given that multisensory neurons can do everything unimodal neurons can, it is legitimate to ask why not all DSC neurons are multisensory or - at least - develop multisensory behavior during an organism's maturation. The novel answer given here derives from an elaboration of a maximum likelihood model of multisensory enhancement by Colonius & Diederich (NIPS, 2002) and Colonius & Diederich (Cognitive, Affective and Behavioral Neuroscience, 2004).

Humans are capable of acting with exquisite temporal precision. In sports such as baseball, athletes can produce movements with a precision of plus or minus 2ms. This ability to precisely time our actions is likely to have provided a major advantage in processes of natural selection. First, I will provide a brief overview from the last three decades of some theoretical accounts of how humans time their actions. Second, I will present some data indicating the task specific nature of visually guided time-to-collision estimation. Finally, I will present some time-to-collision data from a patient with visual form agnosia (patient DF) and some patients with autism.

Most research on the English plural has concentrated on children's acquisition of a rule system (or not!) and on the phonological properties that support generalization. In this talk, I argue that children's understanding of the meaning is also relevant, and I present preliminary data on this issue. I consider three aspects of meaning relevant to the plural -- the number of objects, their similarity to each other, and their status as individuals. The evidence suggests that these three factors matter. Preliminary evidence on parents' use of the plural in talking to their children suggests that the aspects of meaning that children care about are also those that parents care about.

Several years ago Ger-Jan Mensink and I developed a model based on Raaijmakers and Shiffrin’s SAM theory that successfully accounted for the standard data on interference and forgetting. Our model was based on two simple and widely accepted ideas, namely that retrieval from memory is based on associative competition and that context cues affect the activation of information from memory. The same model has also be used to account for spacing effects in human memory.

In recent years however several researchers (e.g. M.C. Anderson) have claimed that interference in memory is not just based on associative competition but that there is also evidence for what they called retrieval inhibition or ‘repression’, the idea that trying to inhibit the retrieval of specific information may cause that information to become less accessible on future retrieval attempts. I will argue that there are alternative explanations for retrieval inhibition that do not rely on the concept of ‘repression’.

Given the recent interest in the cognitive neuroscience of category learning, one question is how existing models from cognitive psychology align with new findings from cognitive neuroscience. In this talk, I present an idea that Brad Love (Texas) and I have developed over the past year which attempts to relate aspects of a clustering model of human category learning to a learning circuit involving the hippocampus, perirhinal, and prefrontal cortex. Results from groups varying in function along this circuit (e.g. infants, amnesics, older adults) are successfully simulated by reducing the model's ability to form new clusters in response to "surprising" events, such as novel stimuli or corrective feedback. It will be suggested that well-specified, process-oriented cognitive models can serve as useful guides for localizing mental function in the brain.

In four experiments we examine whether upright and inverted faces are processed by a single class of neurons with identical response properties, or by two separate and non-interacting populations. Upright and inverted faces were chosen as the stimuli since they reliably show strong N170 responses, and there is a suggestion that upright faces may be processed at least in part using configural mechanisms, which may be supported by a separate population of neurons. In Experiment 1, we examined upright and inverted faces across two contrast levels and one signal to noise ratio, yielding a crossover interaction of the N170 amplitude: when presented in noise, the amplitude of the inverted face was smaller than the upright face, while the reverse is true without noise. In Experiment 2, we showed that the amplitude reversal was robust across all noise levels. In Experiment 3, we varied contrast to see if reversal was a result of degrading a face. We observed no reversal effects. Thus, across conditions, adding noise to full faces was a necessary and sufficient condition for the N170 reversal. This is consistent with a model in which inhibition occurs between neurons processing noise and inverted faces, but not upright faces (configural processing). In Experiment 4, we delayed onset of the upright/inverted face presented in noise. We replicated the smaller N170 for inverted faces at no delay, but observed partial recovery of the N170 for inverted faces at longer delays. These data support a model in which neurons responding to noise inhibit those responding to inverted faces, with inhibition waning to produce selective recovery at longer SOAs. Thus, two visual stimuli processed by the same population may result in a competition for neural representation, which spares stimuli processed by separate populations.

Research reveals that preferences appear dependent on their elicitation method (e.g., choice, pricing) and subject to reversal. Johnson & Busemeyer (2005) have applied Decision Field Theory to these findings and have demonstrated its ability to account for a wide range of preference reversal phenomena. An emergent prediction of DFT's account is that individual-elicited valuations of a gamble will converge to a single amount, regardless of the elicitation method. This prediction contrasts with those made by competing accounts. Further, the model predicts that the variance in response distributions are based on gamble variance, an aspect of preference elicitation ignored by other models.

In this talk I will review the literature on preference reversals and briefly describe alternate models of the phenomena. Next, I will detail DFT extensions to account for preference reversals. Lastly, I will report the results from a new experiment which examines the evolution of preferences across time. The results support the theory, indicating that valuations of a gamble using different response procedures do converge across time, and that response distributions are based on gamble variance, two predictions that deterministic models cannot make.