An Attractor Neural Network Model of Place Field Distortion in the
Rodent Hippocampus

Abstract: We propose to develop a neural network model to test a
popular theory of brain function known as the attractor hypothesis.
The model will reproduce data on neuron firing behavior in rats who
experience changes in the location of visual landmarks.  The work
involves computer simulation of large neural networks, and analysis of
neuron firing data from the lab of Robert Muller, who is collaborating
with the PI, Dave Touretzky.

Research Question and Significance:

The hippocampus is a prominent brain structure thought to be involved
in learning and memory.  In rodents, hippocampal pyramidal cells are
called "place cells" because they fire when the animal is in a
specific place in the environment.  The region of the environment in
which a cell exhibits an elevated firing rate is called the "place
field" of the cell.  Place cell activity is in part a function of
visual landamarks; this has been demonstrated by moving the landmarks
and observing the effect on place field location.

Fenton and Muller (submitted manuscripts) recently showed that place
fields can shrink or stretch when two distinct landmarks are moved
closer together or further apart.  Their landmarks were one black and
one white cue card affixed to the wall of a cylindrical arena.  They
also developed a vector field equation to describe the place field
distortion that results from moving the landmarks.  Although the
equation fits the data well, it does not provide a neurally plausible
explanation for hippocampal processing.

The attractor theory of hippocampus explains hippocampal activity as a
stable state of a dynamical system, implemented in a recurrent neural
network (i.e., a network with feedback connections, similar to the
feedback structure found in hippocampal area CA3.)  If we imagine
place cells spread out over a 2D sheet, a stable state looks like a
"bump" of activity over a local region of the sheet.  This bump is
stabilized by a tendency for cells to excite their neighbors and
inhibit cells further away.

Attractor models are widely used in computational neuroscience, but
realistic attractor models of hippocampus have been difficult to
simulate due to the large numbers of units and connections required.

In the proposed research, we will use an attractor model to simulate
place cells with fields distributed throughout a cylindrical arena.  We
will also develop a representation for visual landmarks, and allow
these visual feature units to project to the hippocampal cells, so
that the cue cards can control place cell firing as they are known to
do in rodents.  Then, we will slide the simulated cue cards closer
together or further apart and measure the distortion in the place
fields.  The goal is to see if the model can be made to agree with
Fenton and Muller's observations.  Their actual data has many
interesting features the model will need to account for, such as the
fact that place fields located close to the cue cards show greater
distortion than those far from the cards, but if one of the cards is
removed entirely, even the most distant fields move in sync with the
remaining card.

Reproducing the Fenton and Muller data is an important test for the
attractor hypothesis.  This particular experiment has not been modeled
before.  In addition, the attractor simulation is on a larger scale
than has been attempted before in the Touretzky lab.  The proposed
model will contain tens of millions of connections.

Project Design and Feasibility

The simulation will be written in Matlab with inner loops coded in C.
It will run on a new, high speed computing facility at the Center for
the Neural Basis of Cognition in the Mellon Institute.  (The facility
consists of a cluster of Pentium III processors with 1 gigabyte of
main memory.)  Matlab will be used for graphical display and analysis
of the simulation results.  

This work involves a collaborative relationship with the Muller lab at
SUNY Brooklyn.  The PI, Dave Touretzky, has already obtained the
neural firing rate data from Bob Muller that we will be attempting to
model.  We will be communicating with the Muller lab throughout the
course of the project.  In addition, Touretzky collaborates closely
with William Skaggs, a neurocientist at the university of Pittsburgh
who does hippocampal physiology and has an interest in modeling.
Touretzky and Skaggs, and their respective students, will all be
interacting during the course of this project.