A Toolkit for Developing Neural Network Models of how the Brain Detects and Orients Attention toward Threat.

sciencediscussionAI and Robotics

Oct 20, 2013 (3 years and 9 months ago)

99 views

Class of 2010, Department of Mathematics and Co
mpu
ter Science, Honors Program,

Mentor: Dr. Robert Dowman
, Department of Psychology.



A Toolkit for Developing Neural

Network Models of

how
the

Brain Detects

and Orients

Attention toward Threat.

Keegan Lowenstein
1
, Dr. Robert Dowman
2

Departments of Computer Science
1

and Psychology
2


C
onnection
ist neural network models have provided important insights into the functional interactions
between the brain areas involved in detecting and orienting attention towards threats to the body. Part of
this
modeling
involves comparing how well different physi
ologically feasible interactions (architectures)
fit the experimental data. Fully exploring these different architectures requires that
many model
parameters be varied.
Varying these param
eters manually
, as has been done in the past,

is unrealistic

and
cre
ates the need for an automated method of model comparison.

For example,
p
revious
architecture
comparisons were
constrained by the use of

fixed connection strength parameters
and were only fit to
behavioral

data and not the brain activity

data
.
In this work
, a

generalized connectionist neural network
model
development and
comparison toolkit

was
implemented

to address these issues
.

The toolkit exists as
a collection

of MATLAB scripts which allow
a
user to define, edit, parameterize, and compare neural
network

models simulating psychological phenomena.

With this toolkit, m
odel comparison is made more
accurate by considering each model architecture using its corresponding optimal connection strength
parameters
.

Having developed this to
olkit, we will
reevaluat
e
t
he model architecture
s

examined in
previous work
, this time
finding the
optimal connection strengths
for each and quantifying how well each
fits
the experimental behavioral and brain
data.