Twenty-Sixth Annual Interdisciplinary Conference
Teton Village, Jackson Hole, Wyoming
January 21 -- 26, 2001
Organizer: George Sperling, University of California, Irvine
ABSTRACTS
===============================================================================
Jeff Mulligan
NASA Ames Research Center
Scan-path Analysis of Air Traffic Control Displays
Air traffic displays are dense with information, but it is not known exactly
which pieces of information are accessed at what times. Better understanding of
how these displays are used, leading to a model of the human operator, will help
display designers to create more effective and efficient displays. To this end,
we investigated visual search performance using simulated air traffic displays.
Subjects viewed displays containing 16 or 20 aircraft including a "conflict
pair," a pair of aircraft on a collision course. Subjects were asked to locate
and identify the conflict pair. All aircraft were at one of four fixed
altitudes. Two methods for representing altitude were investigated: 1) a
numeric string in a small block of text known as the "data tag" (similar to
current air traffic displays); 2) use of a different color to represent the
craft at each of the four altitudes. Eye fixation scan-paths were monitored
using a head-mounted video camera system, a digital video recording system, and
image analysis software. Statistics of empirical scan-path data were compared
with those of synthetic scan-paths generated by a computer model. The computer
model reproduced basic data features by the manipulation of parameters such as
memory decay and the size of the region about fixation within which aircraft
headings can be perceived. We found marked differences in the scan-path data
between conditions with and without color-coding. Without color-coding, we saw
large numbers of very small saccades, produced when the subject read information
in a data tag. The small saccades were, for the most part, absent when color
coding was used; instead, we saw a larger proportion of relatively large
saccades. In a more detailed analysis, we classified each saccade on the basis
of the initial and final fixated objects. For each fixation, the nearest
display object was identified and its type (aircraft icon or data tag, altitude,
etc.) was noted. Each saccade was then placed into a category, such as
craft-to-own-tag, craft-to-same-altitude-craft, etc. When color coding was
employed, subjects tended to scan the aircraft within an altitude group,
resulting in a greater proportion of saccades between objects at the same
altitude, and a larger average saccade size. Subjects used color coding to
efficiently locate and identify aircraft on a collision course. We hope to
generalize the model to allow prediction of performance with operational air
traffic displays.
===============================================================================
Holly Jimison
HRC and OHSU
Challenges in Evaluating the Effectiveness of Web-Based Interventions for
Health Care
Academic time frames for evaluation are no longer acceptable (or of interest)
in providing feedback on rapidly developing technologies. However, there are
many biases associated with information collected from convenient and naturally
occurring experiments in the use of web-based systems.
I will discuss the various techniques that are being used to measure the
effectiveness of web-based approaches to delivering health care information and
decision assistance to patients. The current challenges include
* a rapidly changing background environment of competing web sites and growing
use of the Internet by consumers,
* interventions that are modified and improved frequently over the course of
the trial, and
* a need for rapid feedback.
I will also describe a decision-focused framework for evaluation that makes use
of early feedback on system use.
===============================================================================
Leonid Kontsevich
Smith-Kettlewell Eye Research Institute
How Random Are We?
I will describe a novel approach for finding regularities in time sequences.
This approach has minimum assumptions regarding the regularities to be found
and efficiently uses limited number of examples for finding solution. Applied
to random binary sequences generated by humans it predicts about 80% of
decisions. This algorithm may resemble an associative mechanism for human
memory.
===============================================================================
Richard Shiffrin
Indiana University
Paradoxes of Backwards Induction
I explore the normative bases for reasoning, in the domain of backward
induction as analyzed in two person alternate play games. In the 'centipede'
game of 20 trials, the two players alternate turns, each trying to maximize gain
(not to beat the opponent). A player can end the game, or continue. A decision
to continue loses that player $1 (to the bank) and gives the opponent $10.
Forward reasoning suggests extended play, as each player accumulates much money
by so doing. Backward reasoning suggests stopping on trial one, because it is
always certain the next player will stop (consider trial 20, on which the player
will stop to prevent a sure loss of $1). I propose a resolution to this paradox
that some researchers find strange.
================================================================================
David Heeger
Stanford University
Activity in Primary Visual Cortex Predicts Performance in a Visual Detection
Task
Visual attention can affect both neural activity and behavior in humans. To
quantify possible links between the two, we measured activity in early visual
cortex (V1, V2 and V3) during a challenging pattern detection task. Activity
was dominated by a large response that was independent of the presence or
absence of the stimulus pattern. The measured activity quantitatively predicted
the subject's pattern detection performance: when activity was greater, the
subject was more likely to correctly discern the presence or absence of the
pattern. This stimulus independent activity had several characteristics of
visual attention, suggesting that attentional mechanisms modulate activity in
early visual cortex, and that this attention related activity strongly
influences performance.
================================================================================
Wilson Geisler
University of Texas at Austin
Spatial Coding of Motion Direction in Primary Visual Cortex (tentative)
When an image feature moves with sufficient speed it should become smeared
across space, due to temporal integration in the visual system, effectively
creating a spatial motion pattern that is oriented in the direction of the
motion. Recent psychophysical evidence shows that such "motion streak signals"
exist in the human visual system. Here we report neurophysiological evidence
that these motion streak signals also exist in the primary visual cortex of cat
and monkey. Single neuron responses were recorded for two kinds of moving
stimuli: single spots presented at different velocities and drifting plaid
patterns presented at different spatial and temporal frequencies. Measurements
were made for motion perpendicular to the spatial orientation of the receptive
field ("perpendicular motion") and for motion parallel to the spatial
orientation of the receptive field ("parallel motion"). For moving spot
stimuli, as the speed increases, the ratio of the responses to parallel versus
perpendicular motion increases, and above some critical speed, the response to
parallel motion exceeds the response to perpendicular motion. For moving plaid
patterns, the average temporal tuning function is approximately the same for
both parallel motion and perpendicular motion; in contrast, the spatial tuning
function is quite different for parallel motion and perpendicular motion (band
pass for the former and low pass for the latter). In general, the responses to
spots and plaids are consistent with the conventional model of cortical neurons
with one rather surprising exception: Many cortical neurons are significantly
direction selective for parallel motion. We propose a simple explanation for
"parallel motion direction selectivity" and discuss its implications for the
motion streak hypothesis. Taken as a whole, we find that the measured response
properties of cortical neurons to moving spot and plaid patterns agree with the
the recent psychophysics and support the hypothesis that motion streak signals
are present in V1.
================================================================================
Tatiana Pasternak
University of Rochester
Direction of Motion Is Remembered in Retinotopic Coordinates
We asked if the processing of visual motion information in a visual working
memory task is retinotopicaly organized. Monkeys compared the directions of
two moving random-dot stimuli, sample and test, separated by a temporal delay,
and reported whether the stimuli moved in the same or in different directions.
By presenting the two comparison stimuli in separate locations in the visual
field, we determined whether the storage and retrieval/comparison components of
the task were carried out in retinotopic coordinates. Two psychophysical
measures of direction discrimination provided nearly identical estimates of the
critical spatial separation between sample and test stimuli that lead to a loss
in threshold. Direction range thresholds measured with dot stimuli consisting
of a range of local directional vectors, were affected by spatial separation
when a random-motion mask was introduced during the delay into the location of
the upcoming test. The selective masking at the test location suggests that
the information about the remembered direction was localized and available at
that location. Direction difference thresholds, measured with coherently moving
random dots, were also affected by separation between the two comparison
stimuli. The separation at which performance was affected in both tasks
increased with retinal eccentricity in parallel with the increase in receptive
field size in neurons in cortical area MT. The loss with transfer of visual
information between different spatial locations suggests a contribution of
cortical areas with localized receptive fields to the performance of the memory
task. The similarity in the spatial scale of the storage mechanism derived
psychophysically and the receptive field size of neurons in area MT suggest
that MT neurons are central to this task.
================================================================================
Benjamin T. Backus
University of Pennsylvania
Loss of Disparity Information During Perception of Surfaces
The apparent slant of a stereoscopically defined surface can be altered by
manipulating horizontal disparities, vertical disparities or felt eye position
(e.g. Banks & Backus, 1998). Here we report on physically different stimuli,
containing different patterns of disparity, that are perceptually
indistinguishable from one another. Thus, they are true perceptual metamers
(Loftus & Ruthruff, 1994). Whereas color metamerization occurs at the front
end of the visual system, with the loss of spectral information during the
transduction of light, stereoscopic slant metamerization occurs much later,
with the loss of disparity information during combination with eye position
signals. These metameric stimuli can be made distinguishable by a suitable
change in eye position, as predicted by theory.
==============================================================================
Jan Droesler
Universitaet Regensburg
What Warps Binocular Visual Geometry Away From the Euclidean?
We all were taught in high-school that the sum of angles in a triangle is 180
deg., that the area of a circle is Pi r^2 and some other information about plane
geometry. In the 19th century, however, doubts had arisen as to whether this
information pertains to stringent laws. At the beginning of the 20th century,
Einstein demonstrated that the information was not valid for the physical world
around us in the large. Some 50 years later, Luneburg showed that all these
geometric attributes get changed in a systematic manner, as soon as our spatial
vision is based on binocular visual cues alone: The sum of angles in a triangle
is always smaller than 180 deg., circles have a bigger area than Pi r^2 etc.
The change from monocular to binocular vision induces a change from Euclidean to
non-Euclidean hyperbolic geometry. The cause for this transition has remained
unknown so far. The present paper works out an answer by analyzing the data in
terms of visual automorphisms. There are stimulus transformations, which leave
visual structural properties invariant. Results are that Euclidean automorphisms
can be identified in the case of monocular vision, non-Euclidean hyperbolic
automorphisms in the binocular case.
================================================================================
Steven K. Shevell
University of Chicago
Spatial Localization of a Virtual Object
A remarkable property of human vision is the ability to discriminate the
spatial positions of objects. For example, two abutting vertical lines are
seen to be offset horizontally when their positions differ by just 5 sec of arc
(less than 1/4 inch offset across the length of a football field). Several
theoretical accounts have been offered for spatial localization, including an
ideal-observer model (Geisler, 1984), an aggregation of retinal-location 'signs'
(e.g. the centroid model of Hess & Holliday, 1996), and a contrast-sensitive
spatial-filter model (Wilson, 1986). These models were considered by studying
spatial localization of a perceived object-from-motion: a sphere rotating
around the vertical axis. Measurements using the perceived sphere, which
resulted from successively presented frames of dots, were compared to results
with a stationary pattern of dots (one frame from the sphere) or with randomly
moving dots, which differed from the sphere only with respect to the correlation
between dots' locations in successive frames. The measurements, which showed
the classical drop of positional acuity with contrast, demonstrated best spatial
localization for the perceived sphere (object-from-motion). This cannot be
explained by models that depend on only independent retinal information in each
frame, including ideal observer and centroid models. A neural representation
of inferred shape or contour is consistent with the results (cf. McKee, 1991).
===============================================================================
Joetta Gobell
University of California, Irvine
Effect of Scene Orientation on Depth Perception: Trapezoids, Windsurfers,
Runways
Since Ames's (1951) observation of illusory oscillation of a rotating
trapezoidal figure, numerous factors that affect its depth organization have
been studied. Unlike Ames's rotating trapazoids, our trapazoids merely
oscillate in narrow ranges, and their perceived depth orientation is at issue.
We investigate some traditional factors plus location of axis of rotation,
and a global rotation of the entire scene by 90 deg (see Fig.): (a) Windsurfer
configuration, (b) runway configuration. Method: The stimulus consisted of
two trapezoidal figures and a central fixation point. The arrows indicate
locations of possible axes of rotation. On each trial the trapezoids appeared
as pictured and immediately began rotating through +/- 40 degrees. Orientation
of the trapezoids (e.g. long sides to the left/right or up/down), and an axis
of rotation were randomly assigned on each trial. Observers reported which of
the the two parallel sides appeared to be "in front." Results: Observers
experience qualitatively different perceptions in (a) and (b). In Windsurfer
configuration, the long parallel side is seen in front with the same
probability p whether it appears on the left or the right; p depends on the
observer's sensitivity to perspective versus motion cues. For runway
configurations, when the long parallel side is down, it nearly always appears
in front. When the long side is up, it is most often perceived in the back
as part of an expanding floor. These perceptions are based on world, not
retinal, coordinates. Conclusions: (1) Observers differ substantially in
the weight given to linear perspective and to motion cues in determining
the perceived 3D depth in ambiguous Windsurfer stimuli. (2) Runway stimuli,
which differ from Windsurfers only in their global orientation, give rise to
qualitatively different percepts, indicating the important involvement of
high-level mechanisms in the resolution of these ambiguities.
================================================================================
G.B. Henning, F. A. Wichmann, and C. M. Bird
The Sensory Research Unit, Department of Experimental Psychology
South Parks Road, Oxford
The Pedestal Effect with a Pulse Train and its Constituent Sinusoids
Curves showing threshold contrast for detecting a signal grating as a function
of the contrast of a masking grating of the same orientation, spatial frequency,
and phase show a characteristic improvement in performance at masker contrasts
near the contrast threshold of the unmasked signal. Depending on the percentage
of correct responses used to define the threshold, the best performance can be
as much as a factor of three better than the unmasked threshold obtained in the
absence of any masking grating. The result is called the pedestal effect
(sometimes, the dipper function). We used a 2AFC procedure to measure the
effect with harmonically related sinusoids ranging from 2 to 16 c/deg - all with
maskers of the same orientation, spatial frequency and phase - and with masker
contrasts ranging from 0 to 50%. The curves for different spatial frequencies
are identical if both the vertical axis (showing the threshold signal contrast)
and the horizontal axis (showing the masker contrast) are scaled by the
threshold contrast of the signal obtained with no masker. Further, a pulse
train with a fundamental frequency of 2 c/deg produces a curve that is
indistinguishable from that of a 2-c/deg sinusoid despite the fact that, at
higher masker contrasts, the pulse train contains at least 8 components all of
them equally detectable. The effect of adding 1-D spatial noise is also
discussed.
================================================================================
Joongnam Yang
University of Chicago
Illuminant estimation in surface color perception
When we view an object, it is easy to judge the color of its surfaces,
even though the color signal arriving at the eye has two components: surface
spectral reflectance and spectral illumination. The visual system somehow
disentangles the two so that the appearance of the object is nearly
constant under changes of illumination. This requires that the visual
system somehow discount the illuminant, which can be achieved with knowledge
about the illuminant. The scene contains information about the illuminant
in what are called illuminant cues, which include shadows, specular
reflections, and mutual inter-reflections. One of the cues, specular
reflection, was investigated to determine the role of this cue in color
perception of a scene with 3-D objects and flat surfaces. Two different
methods, cue perturbation and cue elimination, were used. In the
cue perturbation method, all cues in the scene were consistent with one
illuminant except the cue in question, which was consistent with a different
illuminant. The results showed that the specularity cue, when perturbed,
affected color perception, but only when the cue was perturbed from
Illuminant A to D65, not in the other direction. In the cue elimination
method, the specularity cue was entirely removed from the scene. Surface
color perception was affected when the cue was eliminated. Together
these results indicate that specular reflection is an important cue that
affects surface color perception.
==============================================================================
Sophie Wuerger
Keele University
The Intrinsic Blur of the Visual System for Luminance and Chromatic Stimuli
The responsiveness of the human visual system to an image depends on a
multitude of image features, such as the wavelength (colour) of the visual
stimulus and its spatial content. Three main factors limit the spatio-chromatic
sensitivity of the visual system: the optics of the eye, retinal sampling, and
post-receptoral neuronal factors. We investigated a specific aspect of the
spatio-chromatic sensitivity of the human visual system, namely how much blur
the visual system can tolerate in different colour directions and its dependence
on contrast.
Using the model proposed by Watt & Morgan (1983 Vis. Res., 23, 1457-1477) we
estimated the internal blur for each colour direction and arrived at the
following estimates: 1 minute of visual angle for red-green and luminance, and
1.8 minutes of visual angle for yellow-blue. Furthermore, the contrast
dependence of blur sensitivity is identical for red-green and luminance.
We conclude that for (stationary) stimuli the blur tolerance in the luminance
and the red-green channel is predicted by the absolute cone contrast and is
independent of the sign of the L and M cone contrast. (luminance: L and M cone
contrast of same sign; red-green: L and M cone contrast of opposite sign). Our
results are consistent with the hypothesis that stationary luminance and
red-green stimuli are encoded by the same mechanism.
Blur tolerance is not predicted by the known contrast sensitivity function for
luminance, red-green, and yellow-blue gratings. Our measurements of the
chromatic blur tolerance of the human visual system are potentially useful for
image processing when when lowpass filters are used for noise removal.
===============================================================================
Zygmunt Pizlo
Purdue University
Status of the Zone Theory of Color Coding
The so called zone theory incorporates the trichromatic theory and the
opponent process theory. The trichromatic and opponent process theories make
contradictory predictions, which poses a logical problem for the zone theory.
These three theories will be discussed in the context of experiments by
Helmholtz (1852), Maxwell (1856), Hecht (1928) and Hurvich & Jameson (1951).
================================================================================
David J. Fleet
Xerox PARC and Queen's University
Bayesian Detection and Tracking of Motion Boundaries
Visual motion at occlusions is a rich source of information about the location
of surface boundaries and about the depth ordering of surfaces at these
locations. Despite this, models for the processing of motion boundaries in
biological systems are rare. In machine vision, these "motion boundaries" are
most often viewed as a source of noise for current motion estimation techniques
which assume motion is smooth.
We propose a Bayesian framework for representing and estimating image motion
in terms of multiple motion models, including both smooth motion and local
motion discontinuity models. We compute the posterior probability distribution
over models and model parameters, given the image data, using discrete samples
and a particle filter for propagating beliefs through time. This talk will
introduce the problem and describe our Bayesian approach, including our
generative models, the likelihood computation, the particle filter, and a
mixture model prior from which samples are drawn.
================================================================================
Tjeerd Dijkstra
Ohio State University
Perception of Orientation: an Empirical Bayesian Model
Perception of the orientation of objects is important in our
interaction with the environment. So far, research has focused on the
fronto-parallel orientation of lines and gratings with the main
result that vertical and horizontal orientations are perceived more
accurately and precisely than oblique ones (oblique effect).
We tested the orientation perception of fronto-parallel ellipses with
different length-to-width (aspect) ratios in various orientations. A
circle was included in the test set. Six naive subjects adjusted a
broken line (probe) to match the major axis orientation of an ellipse
that was placed at the center of the probe.
The precision of the settings as quantified by the circular standard
deviation (CSD), increased with decreasing aspect ratio.
Reparametrizing aspect ratio as index of difficulty (defined as the
inverse of aspect ratio minus one), CSD increased linearly with index
of difficulty. This result could be captured by a simple ideal
observer model were the vertices of the polygon making up the ellipse
were perturbed with noise: a single noise level for each subject was
sufficient to capture the results.
The accuracy results show large biases, especially for the low aspect ratios
(close to a circle). For the circle, subjects had a non-uniform distribution of
settings. Furthermore, there are large individual differences among the
subjects. We can explain these differences by a Bayesian model that takes the
distribution of settings to the circle as a prior distribution. Thus the prior
is obtained from the settings to a neutral stimulus. Going beyond the domain of
perception of orientation, we believe empirical Bayesian modeling to be a useful
new tool for vision research.
================================================================================
Misha Pavel
Oregon Granduate Institute
Fusion-Based Robust Signal Processing by Humans and Machines
In my presentation, I will briefly note the wide range of benefits that can
be derived from biological and machine data fusion, but I will then focus on
fusion in service of pattern recognition. Although many existing automatic
pattern recognition systems have achieved considerable success over the past
fifty years, most of them lack robustness - the ability to perform as well as
possible in unpredictable and changing environments. In contrast, biological
systems seem to be much more resilient to the environmental changes that are,
at least partially, irrelevant to the tasks. The fact that current artificial
systems lack the robustness found in natural systems leads to questions that
address the basic differences between the natural and the statistically optimal
approaches. Our preliminary analysis of these differences led us to hypothesize
new methodology for pattern recognition. I will briefly describe a working
hypothesis whereby data fusion in conjunction with neural-like computation can
be used to achieve more robust pattern recognition performance that that
obtained with more traditional approaches. I will illustrate the approach on
one or two specific and realistic examples.
================================================================================
RECOGNITION MEMORY SESSION
The dominant model of recognition memory has been a single-process
continuous-state model that assumes that memory access consists of the
interaction of retrieval cue and a memory structure (Anderson, 1973; Gillund
& Shiffrin, 1984; Murdock, 1982). The result of memory access is a level of
familiarity associated with a test item, and this serves as the sole source of
information on which to base the recognition decision. While dominant in the
field, few researchers believe that recognition is always based only on the
familiarity of the test item (cf. Gillund & Shiffrin, 1984). It is strongly
suspected, at least at times, that the retrieval of information from memory also
contributes to recognition. That is, a dual-process model of recognition is
tacitly accepted that incorporates both a familiarity-based and a retrieval-
based process (e.g. Atkinson & Juola, 1973; Gillund & Shiffrin, 1984; Yonelinas,
1997). The problems for the field have been to identify empirical phenomena
that require retrieval-based memory access in order to be explained and how to
measure the contributions of the different memory access processes. The purpose
of the proposed AIC session is to explore theories of recognition memory,
measurement issues, and empirical evidence that bears on the dual-process issue.
================================================================================
Rik Henson
Institute of Cognitive Neuroscience & Wellcome Department of Cognitive Neurology
London, England
fMRI Studies of Recognition Memory
Recent functional magnetic resonance imaging (fMRI) studies of recognition
memory for verbal material have revealed a network of prefrontal and parietal
regions associated with successful recognition. I will describe attempts to
dissociate activity in these regions according to the distinction between
recollective vs. nonrecollective recognition, using experimental manipulations
such as the Inclusion/Exclusion procedure, Remember vs. Know judgments and
confidence judgments. Though the results from these different manipulations are
highly consistent, none can be said conclusively to isolate recollective
processes. The relative scarcity of medial temporal activations in these
studies also remains a puzzle. Nonetheless, the studies highlight the role of
decision processes in yes/no recognition memory, which may have behavioural
consequences that have been underemphasised by standard dual-process theories.
===============================================================================
Ken Norman
University of Colorado, Boulder
Modeling Hippocampal and Neocortical Contributions to Recognition Memory
Dual-process models posit that recognition judgments are based on recall of
specific details, and on feelings of familiarity (that index, holistically, how
well the test probe matches stored memory traces). One way to place
dual-process models on stronger footing is to look at how the brain implements
recall and familiarity; in recent years, a very clear story has emerged whereby
recall depends on the hippocampus, and familiarity-based recognition is
supported by the medial temporal neocortical regions that surround the
hippocampus. To explore how these structures contribute to recognition, we
have constructed neural network models that incorporate key anatomical and
physiological properties of the hippocampus and neocortex, respectively --
these models provide a principled way of predicting how manipulations will
affect recall and familiarity. One prediction is that a list strength effect
should be present for hippocampally-driven recall, but not for neocortically-
driven familiarity; I will explain why the models make this prediction, and I
will present new empirical results consistent with this prediction. More
generally, I hope to demonstrate that mathematical models of recognition can
benefit by tapping into our growing knowledge of how the brain gives rise to
recognition memory.
===============================================================================
Elliot Hirshman
University of Colorado, Denver
Pharmacological "Lesions"
===============================================================================
Michael D. Rugg
Institute of Cognitive Neuroscience & Wellcome Department of Cognitive Neurology
London, England
Fractionation of Recognition Memory: Convergent Evidence?
I will briefly review electrophysiological (ERP) studies relevant to the
question whether recognition memory is supported by two (or more) underlying
processes, and then present some new ERP data that both supports and qualifies
previous findings. I will also make some general remarks, in light of the other
contributions to the session, on the extent to which different approaches to
the questions of whether and how recognition memory should be modelled as
multiple processes converge on a common answer.
===============================================================================
William Banks
Pomona College
Multidimensional Analysis of Memory
This talk presents a method for creating a multidimensional representation of
memory and using it to predict recognition memory, source memory, exclusion
performance, and "false fame" effects. Despite the fact that these memory
domains use different paradigms, methods of analysis, and theoretical
interpretations, their results can be treated with a single analytic model. We
show how to generate a multidimensional memory representation based on
signal-detection theory (a version of General Recognition Theory) and make
predictions for each of these paradigms. The detection model is simpler
than comparable multinomial models, it is more easily generalizable, it is much
easier to image and think about, and it does not make threshold assumptions.
Results show clearly and intuitively the relationship between exclusion and
source discrimination and how decision processes in the multidimensional space
can result in effects like false fame. Several other topics, including memory
representations of faces, will be addressed.
==============================================================================
Charles K. Brainerd
University of Arizona
Representational Bases for Dual-Recognition Processes
In fuzzy-trace theory's approach to dual-recognition processes, the tendency
of different types of memory representations (especially, verbatim and gist
traces) to provoke different types of retrieval is stressed. The theory's
assumptions are implemented in the conjoint-recognition model. Results from
experimental applications of this model will be reviewed. The question of
whether recent findings on the model's phantom recollection parameter require
the introduction of a third recognition process will be examined and possible
extensions to recall will be discussed.
==============================================================================
Caren Rotello
University of Massachusetts
ROC Analyses of Recognition Memory and Remember-Know
The Remember/Know paradigm has been used extensively to distinguish between
two subjective states of awareness in recognition: explicit recollection
(remembering) of the event, and more general familiarity (knowing) that suggests
the event occurred. Recently, a single-process familiarity-based model of the
data from that paradign has been presented. That model makes predictions about
the nature of the receiver-operating characteristic (ROC) curve, which I will
discuss and evaluate. Dual-process and multidimensional models of the
Rembmber/Know data will also be considered.
================================================================================
Bosco S. Tjan
University of Southern California
Object Recognition by Anarchy
Most conventional theories of object recognition assume that within a single
visual-processing pathway only one form of representation is derived and used to
recognize objects. Versatility of the visual system comes from having multiple
visual-processing pathways, each supporting a particular class of objects or
recognition tasks. We propose and simulate a theoretically simpler alternative,
capable of explaining the same set of data and more. A single primary
visual-processing pathway, loosely modular, is assumed. Memory modules are
attached to sites along this pathway. Object-identity decision is made
independently at each site. A site's response time is a monotonic-decreasing
function of its confidence regarding its decision. The system's response is the
first-arriving response from any site. The effective representation of such a
system, determined behaviorally, appears to be self-adapting and specialized for
different tasks and stimuli. This is merely a reflection of the decisions being
made at the appropriate level of abstraction.
===============================================================================
Georg Meyer
Keele University
What Holds Speech Together? - Perceptual Organization by Formant Structure
We are remarkably good at understanding speech in background noise. A
possible explanation for this is that we organise our auditory environment into
multiple streams that we can selectively attend to. Experimental data suggests
that simple features, such as the fundamental frequency or the spectro-temporal
structure of sounds, can be used to segregate competing sounds into separate
streams. The segregation process only requires low-level knowledge of the
structure of sound, such as that all harmonics in an auditory scene that share
the same fundamental frequency are likely to be produced by a single source.
The ideas underlying auditory scene analysis and an application of this
segregation strategy to speech recognition in noise will be discussed in the
first part of the talk.
A fundamental problem of auditory scene analysis, when applied to speech
sounds, is the inherent contrastive nature of speech, which means that
successive speech segments of a single speaker are designed to maximally
different. A 'good auditory speech analyser' should separate a single speech
stream into many different fragments. We know that this is not the case because
we perceive speech as a single stream. I will use the second part of the talk
to describe some experiments that suggest that formant structure one of the
cues used to 'hold speech together.'
===============================================================================
Octavio Betancourt, Computer Science and John Antrobus (presenter),
Experimental Cognition, City College or the City University of New York
Using Natural Harmonics as Acoustic Features in Speech Recognition:
A Vowel Classification Example
By exploiting the contextual dependencies in speech category information
across successive data windows, automatic word recognition systems are
able to tolerate substantial error within individual windows. Nevertheless,
the limited accuracy of these systems across large vocabulary and speakers
sets, and with running speech, suggests that any algorithm that can
substantially reduce error in the early stages of recognition may improve
the accuracy of the entire word classification system. We suggest that the
accuracy of these systems can be improved substantially if the analysis of
the periodic portions of speech, namely voiced speech, including both vowels
and consonants, is improved. In order to simplify the search for, and test
of, new algorithms, we have restricted our initial examination to isolated
vowels in an /hVd/ format.
Because automatic recognition systems compute the power spectra of
the speech signal from fixed windows, the Fourier representation includes
error terms introduced by the disparity between the natural period of the
voiced speech, T0 , and the arbitrarily chosen window length. Although the
standard procedure of multiplying the original signal by a window function
equal to zero at both ends results in a periodic function, it nevertheless
produces harmonic distortions that, in isolated vowel recognition,
contribute to speech errors that are 15% higher, than those of human
listeners.
Our method eliminates these errors by computing T0, F0, the fundamental
frequency, and its natural harmonics, within each fixed window, thereby
reducing the error rate of the tradition method, as measured by a Euclidean
classifier, from 20% to 13%. A second algorithm, which transposes the
spectrum proportional to F01/3, minimizes speaker variability in these
spectra, reducing classification error an additional 6.5% to within 1% of
human accuracy. Most remarkably, that this high accuracy is accomplished
by a simple Euclidean norm classifier, and is as high as that achieved by
a high dimensional Bayesian, Maximum Likelihood Discriminant Function
classifier, indicates that our method represents vowels patterns with
great precision and parsimony.
===============================================================================
Barbara Dosher
University of California, Irvine
Using Stimulus Noise to Define Attentional Processes
================================================================================
Philip L. Smith
University of Melbourne
Attentional Dynamics of Masked Detection
The role of spatial attention in the processing of elementary stimulus
attributes has, until recently, been somewhat unclear. One group of studies
supports the view that simple stimulus attributes are detected preattentively
and that focal attention is involved only in higher-order processing, such as
that subserving the identification of complex forms. A second group of studies
supports the view that detection sensitivity is enhanced for stimuli presented
at attended locations. Data reported by Smith (2000) suggest that the variable
which determines whether or not a signal enhancement effect is obtained is
whether or not backwardly-masked stimulus presentation is used.
In this talk I review new and existing evidence for the masking hypothesis.
I then present a model of stimulus processing in cued detection, the
attention-gated stochastic integrator, which predicts the different attentional
effects obtained with masked and unmasked presentation. This model, which
combines the multichannel leaky stochastic integrator model of Smith (1995)
and the episodic theory of attention of Sperling and Weichselgartner (1995),
assumes that the rate of information accrual in a sequential-sampling decision
mechanism is gated by selective attention. Differential predictions for masked
and unmasked presentation follow from the assumption of greater informational
persistence for unmasked stimuli.
===============================================================================
Roger Ratcliff, Mark Segraves, and Anil Cherian
Northwestern University
Neural Recordings and Simple Two-choice Decisions
We report data from a simple two-choice task using rhesus monkeys as subjects.
The behavioral results are presented as accuracy rates and reaction time
distributions for correct and error responses. We recorded from buildup/prelude
cells in the superior colliculus during the task. The diffusion model was fit
to the behavioral data and the estimates of decision time from the model were
shown to match the estimates from the neural recordings. We also found evidence
for competition between the two responses in neural activity: when response A
was highly likely and the monkey made response B, there was increased activity
in the cells corresponding to response A (though activity in the cells
corresponding to B was higher). We discuss the current state of modeling
two-choice tasks in the neurobiological domain.
===============================================================================
George Sperling
University of California, Irvine
TBA
================================================================================
Thomas Busey
Indiana University
Recognition and Confidence Judgments of Faces: Contributions of Recollection
and Familiarity
Using blended stimuli we demonstrate a dissociation between confidence and
recognition performance in experiments with naturalistic faces. Then in a
series of four experiments we rule out explanations for this dissociation based
on a) signal detection theory, b) unequal variance SDT, c) global familiarity
(density) models, and d) sampling models. We argue that the results support a
process in which both recollective and familiarity processes play a role. The
results bear not only on confidence judgments of faces (important for eyewitness
testimony) but also for models of face recognition that rely on a representation
of the similiarity between faces as input.
==============================================================================
Amy Heather Criss
Indiana University
Why are you confused? Item and Context Noise in Recognition Memory
The purpose of the comment is two-fold. First, we point to the close
similarities of BCDMEM, the model proposed by Dennis & Humphreys (in press),
to extant global familiarity models of recognition memory. One of the primary
differences being the order in which item and context information are used as
retrieval cues. Unfortunately, there may be no plausible way to test these
assumptions. Our second goal is to consider alternatives to BCDMEMs strict
assumption that item noise does not contribute to recognition decisions. Item
noise refers to interference from items, other than the test item, on the study
list. Context noise refers to interference from incorrect contexts (e.g. from
presentations of the test item prior to the study list). We argue, and
demonstrate in two experiments, that both sources of noise affect recognition
performance contrary to the strong position of Dennis & Humphreys (in press) in
which item noise plays no role.
===============================================================================
Tim Rickard
University of California, San Diego
Memory Retrieval Practice: Strengthening or Instance Accrual?
A commonly held view in the field is that most learning in long-term memory
involves independent instance, or exemplar, accrual. Logan (1988) extended
this view to account even for RT improvement with practice on specific items.
However, this extreme case appears to constitute the ideal conditions for
prototype learning, or strengthening, if in fact such a process occurs in
nature. To explore this issue, RT distributions fits were evaluated for
simplest case instance and strengthening models of memory retrieval practice.
Two methodological features were: 1) separate distribution fits to the practice
data for each item, and 2) de-convolution of the perceptual-motor component of
the RT distribution based on independent data. The results suggest a
fundamental weakness in the instance distribution model, as well as a somewhat
surprising shape in the de-convolved retrieval distribution. Overall, the
results favor a strengthening account, raising the question of what principle(s)
determines whether practice yields strengthening or instance accrual in a given
task domain.
===============================================================================
David E. Huber
University of Colorado, Boulder
Source Confusion and Discounting in Short-term Word Priming: Feature-based
versus Word-based Accounts
Huber et al. (in press) observed 1) a preference for or against prime-related
words depending upon the manner in which prime words were processed and 2)
performance deficits with priming, regardless of the direction of preference.
In the probabilistic feature-based model, Responding Optimally with Unknown
Sources of Evidence (ROUSE), both the preference for prime-related words as well
as the priming deficits arise from source-confused feature activation. A switch
to a preference against prime-related words is explained in terms of the
optimal discounting of primed features. An alternative instantiation of source
confusion and discounting (Ratcliff and McKoon, in press) equally accounts for
these results applying the same mechanisms at the word-level. New experiments
are presented providing evidence in support of the feature-based approach found
in ROUSE.
===============================================================================
Kenneth J. Malmberg
Indian University
How Study Time Affects Implicit and Explicit Memory: The "One-Shot" Hypothesis
Direct, explicit, memory performance (e.g. recognition and recall) is improved
by both spaced and massed study of an item. However, indirect, implicit, memory
performance (e.g. word fragment completion and perceptual identification) is
improved only by spaced study. Within the framework of the 'Retrieving
Effectively from Memory Theory' (REM), these results are predicted when a new
assumption, called the one-shot hypothesis, is incorporated: Increasing the
number of massed repetitions or the time of massed study increases the
item-content information stored, but (beyond some minimum time or number) does
not increase the amount of context information stored. Experiments using source
memory and free recall procedures verify this one-shot hypothesis, and suggest
that context is fully encoded within approximately two seconds.
===============================================================================
Simon Dennis
University of Queensland
The Syntagmatic Paradigmatic Model of Sentence Processing
While connectionist models of sentence processing (e.g. Simple Recurrent
Network, SRN, Elman 1993; Visitation Set Grammar model, VSG, Tabor &
Tannenhaus 1999) pose a significant challenge to symbolic accounts, they also
have a number of limitations. They are difficult to scale to substantive
portions of a language, in terms of the size of the vocabularies they can
accommodate, the length of the sentences they can process and the number of
grammatical structures they capture. In addition, it has been argued that
connectionist models are not able to account for the systematic nature of
language use (Fodor & Pylyshyn 1988, Marcus 1999, although see Elman 1998 for
counter arguments). Furthermore, it is unclear how models such as the SRN or
the VSG will be able to account for the affects on reading when a sentence is
immediately preceded by syntactically and relationally similar sentences.
The Syntagmatic Paradigmatic (SP) model of sentence processing assumes that
sentences are stored in memory as distributed traces of syntagmatic (between
slot) and paradigmatic (within slot) associations. Sentence interpretation
involves using the current sentence fragment as a cue to memory (employing the
Minerva II model, Hintzman 1984, 1986). The retrieved vector is then treated
as a set of constraints on working memory resolution of the sentence. The SP
model scales well to large (> 145000 sentence) naturally occurring corpi and
demonstrates strong systematicity. In addition, the working memory structures
that contain the retrieved traces contain the residue of previous sentences, so
that the model can account for both syntactic and relational sentence priming.
================================================================================
Mark Steyvers
Stanford University
Small World Networks and Semantic Networks
Watts and Strogatz (1998) showed that many real life networks such as the
electric power grids, social networks, and the neural network of the worm
Caenorhabditis Elegans are small-world networks. These networks exhibit small
average distance between nodes as well as strong local clustering. We show
that several types of semantic networks, e.g. associative networks and networks
formed by linking words with the contexts they appear in are all small world
networks. We will show how to make use of the small average distances in
semantic networks to place words in a low dimensional space. The distances
between words in this space can predict confusions in recognition memory and
recall.
================================================================================
Last update: 30jan01