Recent publications from GURU

[Gary Cottrell]

Hello all,

Below are titles of six recent papers from Gary's
Unbelievable Research Unit (GURU). Five of these will
appear in the 1998 Proceedings of the Cognitive Science
Society. One will appear in the Proceedings of Special
Interest Group on Information Retrieval.

All are available from my home page:

	http://www-cse.ucsd.edu/users/gary/

Abstracts are appended to the end of this message.

Cheers,
gary

Gary Cottrell 619-534-6640 FAX: 619-534-7029 
Faculty Assistant Joy Gorback: 619-534-5948
Computer Science and Engineering 0114
IF USING FED EX INCLUDE THE FOLLOWING LINE:          "Only connect"
3101 Applied Physics and Math Building
University of California San Diego			-E.M. Forster
La Jolla, Ca. 92093-0114

Email: gary at cs.ucsd.edu or gcottrell at ucsd.edu



Anderson, Karen, Milostan, Jeanne C. and Cottrell,  Garrison
W.  (1998)  Assessing  the contribution	of representation to
results.  In Proceedings of the	Twentieth  Annual  Cognitive
Science	Conference, Madison, WI, Mahwah: Lawrence Erlbaum.

Clouse,	Daniel S. and Cottrell,	Garrison W. (1998) Regulari-
ties  in a Random Mapping from Orthography to Semantics.  In
Proceedings  of	 the  Twentieth	 Annual	 Cognitive   Science
Conference, Madison, WI, Mahwah: Lawrence Erlbaum.

Dailey,	Matthew	N., Cottrell, Garrison W. and Busey,  Thomas
A. (1998) Eigenfaces for familiarity.  In Proceedings of the
Twentieth Annual Cognitive Science Conference, Madison,	 WI,
Mahwah:	Lawrence Erlbaum.

Laakso,	Aarre and Cottrell, Garrison W.	 (1998)	 How  can  I
know what You think?:  Assessing representational similarity
in neural systems.  In Proceedings of the  Twentieth  Annual
Cognitive  Science Conference, Madison,	WI, Mahwah: Lawrence
Erlbaum.

Padgett, Curtis	and Cottrell, Garrison W.  (1998)  A  simple
neural	network	 models	 categorical  perception  of  facial
expressions.  In Proceedings of	the Twentieth Annual  Cogni-
tive  Science Conference, Madison, WI, Mahwah: Lawrence	Erl-
baum.

Vogt,  Christopher  C.	and  Cottrell,	Garrison  W.  (1998)
Predicting  the	performance of linearly	combined IR systems.
In Proceedings of  Special  Interest  Group  on	 Information
Retrieval.

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Abstracts

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Assessing the Contribution of Representation to Results
 
Karen Anderson kanders at cs.ucsd.edu
Jeanne Milostan jmilosta at cs.ucsd.edu
Garrison W. Cottrell gary at cs.ucsd.edu
Computer Science and Engineering Department 0114
Institute for Neural Computation
University of California San Diego
La Jolla, CA 92093-0114 


In this paper, we make a methodological point concerning
the contribution of the representation of the output of a
neural network model when using the model to compare to
human error performance.  We replicate part of Dell,
Juliano \& Govindjee's work on modeling speech errors
using recurrent networks (Dell et al. 1993).  We find
that 1) the error patterns reported by Dell et al. do not
appear to remain when more networks are used; and 2) some
components of the error patterns that are found can be
accounted for by simply adding Gaussian noise to the
output representation they used. We suggest that when
modeling error behavior, the technique of adding noise to
the output representation of a network should be used as a
control to assess to what degree errors may be attributed
to the underlying network.

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Regularities in a Random Mapping from Orthography to Semantics

Daniel S. Clouse and Garrison W. Cottrell
Computer Science & Engineering 0114
University of California, San Diego
La Jolla, CA  92093

{dclouse,gary}@cs.ucsd.edu

In this paper we investigate representational and
methodological issues in a attractor network model of the
mapping from orthography to semantics based on (Plaut,
1995).  We find that, contrary to psycholinguistic
studies, the response time to concrete words (represented
by more 1 bits in the output pattern) is slower than for
abstract words.  This model also predicts that response
times to words in a dense semantic neighborhood will be
faster than words which have few semantically similar
neighbors in the language.  This is conceptually
consistent with the neighborhood effect seen in the
mapping from orthography to phonology (Seidenberg &
McClelland, 1989; Plaut et al. 1996) in that patterns with
many neighbors are faster in both pathways, but since
there is no regularity in the random mapping used here, it
is clear that the cause of this effect is different than
that of previous experiments.  We also report a rather
distressing finding.  Reaction time in this model is
measured by the time it takes the network to settle after
being presented with a new input.  When the criterion used
to determine when the network is ``settled'' is changed to
include testing of the hidden units, each of the results
reported above change the direction of effect -- abstract
words are now slower, as are words in dense semantic
neighborhoods.  Since there are independent reasons to
exclude hidden units from the stopping criterion, and this
is what is done in common practice, we believe this
phenomenon to be of interest mostly to neural network
practitioners.  However, it does provide some insight into
the interaction between the hidden and output units during
settling.

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Eigenfaces for Familiarity
 
Matthew N. Dailey mdailey at cs.ucsd.edu
Garrison W. Cottrell  gary at cs.ucsd.edu

Computer Science and Engineering Department 
University of California, San Diego 
9500 Gilman Dr., La Jolla CA 92093-0114 USA  

Thomas A. Busey busey at indiana.edu
Department of Psychology 
Indiana University 
Bloomington, IN 47405 USA

A previous experiment tested subjects' new/old judgments
of previously-studied faces, distractors, and morphs
between pairs of studied parents.  We examine the extent
to which models based on principal component analysis
(eigenfaces) can predict human recognition of studied
faces and false alarms to the distractors and morphs.  We
also compare eigenface models to the predictions of
previous models based on the positions of faces in a
multidimensional ``face space'' derived from a
multidimensional scaling (MDS) of human similarity
ratings.  We find that the error in reconstructing a test
face from its position in an ``eigenface space'' provides
a good overall prediction of human familiarity ratings.
However, the model has difficulty accounting for the fact
that humans false alarm to morphs with similar parents
more frequently than they false alarm to morphs with
dissimilar parents.  We ascribe this to the limitations of
the simple reconstruction error-based model.  We then
outline preliminary work to improve the fine-grained fit
within the eigenface-based modeling framework, and discuss
the results' implications for exemplar- and face
space-based models of face processing.


XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

How Can *I* Know What *You* Think?:
Assessing Representational Similarity in Neural Systems

Aarre Laakso aarre at ucsd.edu
Department of Philosophy
University of California, San Diego
La Jolla, CA 92093

Garrison W. Cottrell gary at cs.ucsd.edu
Institute for Neural Computation
Computer Science and Engineering
University of California, San Diego
La Jolla, CA 92093

How do my mental states compare to yours? We suggest that,
while we may not be able to compare experiences, we can
compare neural representations, and that the correct way
to compare neural representations is through analysis of
the distances between them.  In this paper, we present a
technique for measuring the similarities between
representations at various layers of neural networks.  We
then use the measure to demonstrate empirically that
different artificial neural networks trained by
backpropagation on the same categorization task, even with
different representational encodings of the input patterns
and different numbers of hidden units, reach states in
which representations at the hidden units are similar.


XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

A Simple Neural Network Models Categorical
Perception of Facial Expressions

Curtis Padgett and Garrison W. Cottrell
Computer Science & Engineering 0114
University of California, San Diego 
La Jolla, CA  92093-0114

{cpadgett,gary}@cs.ucsd.edu


The performance of a neural network that categorizes
facial expressions is compared with human subjects over a
set of experiments using interpolated imagery. The
experiments for both the human subjects and neural
networks make use of interpolations of facial expressions
from the Pictures of Facial Affect Database (Ekman &
Freisen, 1976).  The only difference in materials between
those used in the human subjects experiments (Young et
al., 1997) and our materials are the manner in which the
interpolated images are constructed -- image-quality
morphs versus pixel averages.  Nevertheless, the neural
network accurately captures the categorical nature of the
human responses, showing sharp transitions in labeling of
images along the interpolated sequence.  Crucially for a
demonstration of categorical perception (Harnad, 1987),
the model shows the highest discrimination between
transition images at the crossover point. The model also
captures the shape of the reaction time curves of the
human subjects along the sequences. Finally, the network
matches human subjects' judgements of which expressions
are being mixed in the images. The main failing of the
model is that there are intrusions of ``neutral''
responses in some transitions, which are not seen in the
human subjects. We attribute this difference to the
difference between the pixel average stimuli and the image
quality morph stimuli.  These results show that a simple
neural network classifier, with no access to the
biological constraints that are presumably imposed on the
human emotion processor, and whose only access to the
surrounding culture is the category labels placed by
American subjects on the facial expressions, can
nevertheless simulate fairly well the human responses to
emotional expressions.

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Predicting the Performance of Linearly Combined IR Systems 

Christopher C. Vogt
University of California, San Diego, CSE 0114, La Jolla, CA 92093, USA 

Garrison W. Cottrell
University of California, San Diego, CSE 0114, La Jolla, CA 92093, USA 

Abstract

We introduce a new technique for analyzing combination
models. The technique allows us to make qualitative
conclusions about which IR systems should be combined. We
achieve this by using a linear regression to accurately
(r2=0.98) predict the performance of the combined system
based on quantitative measurements of individual component
systems taken from TREC5. When applied to a linear model
(weighted sum of relevance scores), the technique supports
several previously suggested hypotheses: one should
maximize both the individual systems' performances and the
overlap of relevant documents between systems, while
minimizing the overlap of nonrelevant documents. It also
suggests new conclusions: both systems should distribute
scores similarly, but not rank relevant documents
similarly. It furthermore suggests that the linear model
is only able to exploit a fraction of the benefit possible
from combination. The technique is general in nature and
capable of pointing out the strengths and weaknesses of
any given combination approach.

SIGIR'98
24-28 August 1998
Melbourne, Australia.