Transcript No Slide Title

MODELING THE RECPTIVE FIELD ORGANIZATION OF
OPTIC FLOW SELECTIVE MST NEURONS
+
Yu ,
+
Gaborski ,
Chen-Ping
William K. Page*, Roger
and Charles J. Duffy
Dept. of Neurology, Univ. of Rochester, Rochester, NY 14642
+Dept. of Computer Science, Rochester Institute of Technology, Rochester, NY 14623
The local motion model of 819R09
shows an irregular fit to the optic flow
response data, suggesting local
motion mechanisms partially account
for the global pattern selectivity.
METHODS: MST Neuronal Responses to Optic Flow and Local Motion
80
Simulate Observer Movement in 16
Directions
60
Inhibitory
Gaussian
Parameters
Length=Gain
Head=Width
Model Testing Fit
Model Training Fit
Neuron
Model
Neuron
Local Motion Stimuli
Model
Optic Flow Stimuli
Dual Simultaneous Stimulation
Single Site Stimulation
Two Hot Spot Directions X 4 Test Spot Directions
One Site X 4 Directions
(4 directions of local motion at 9 sites 30o2)
Hot
Spot
50
50
spks/s
50
500 ms
Test
Spot
Test
Spot
spks/s
20
500 ms
500 ms
500 ms
0
Neuron
819R10
Neuron 819R09
25 models with
fewest response
group error
(3 clusters)
00111001….. X 00111001…..
Cross-over models at random sites to yield 2550 new models
Repeat across 75 generations (asymptotic error reduction)
Dual Stimulus Model of Optic Flow Responses
Class Error: 15
Model
Neuron
Optic Flow Stimuli
Class Error: 5
Model
Neuron
Optic Flow Stimuli
Neuron 819R34
PRELIMINARY SUMMARY
CONTINUED DEVELOPMENT
Excitatory
Gaussian
Vector differences between the Single Site Data
single and the dual site data (Only Dual Sites Shown)
represent dual site interactions
with that Hot Spot direction.
• MST neuronal responses to optic
flow are not accounted for by the array
of local motion responses.
• Evaluate model across sample of 60
neurons recorded with optic flow, single
site, and selected dual site stimuli.
Inhibitory
Gaussian
Transforms for sites not in dual
site study are then interpolated
from neighboring sites.
• Dual Gaussian models derived by
genetic algorithm fit single site local
motion, but not optic flow responses.
• Assess impact of the dual site transforms
in modeling early phasic responses versus
late tonic responses to local motion and
optic flow stimuli.
9 Site, Dual Gaussian Model
Of MST Receptive Fields
Assess fit of each model to neuron response data
25 models with
least total error
across stimuli
(firing rate)
Neuron 819R34
METHODS: 2 Site Data Changes 1 Site Model of Optic Flow Response
METHODS: Dual Gaussian Response Field Modeling
Each model: 18 Gaussians (2 for each of 9 sites)
Gain
Direction Width Polarity
0011100 101110101 1100101 001101100
Center Hot Spot Right
Center Hot Spot Down
Responses to optic flow were
predicted by the version of the
model having the local motion
direction at the tested Hot Spot.
Single Stimulus Model of Optic Flow Responses
50
Randomly Generate 2550 Gaussian Models
Singles Model
We compared single and dual stimulus models by ability to fit optic flow responses.
Responses were divided in to three levels by k-means cluster analysis (typically
either: no / small / large response or inhibitory / no / excitatory response). The diverse set of
results is assessed by the number of points that matched cluster classification.
40
Training with Genetic Algorithm
Center Hot Spot Left
We hypothesized that interactions between local response mechanisms might alter
the net directionality of MST receptive fields and promote global pattern selectivity.
We tested this hypothesis by simultaneously presenting local motion stimuli at two
sites in the receptive field, revealing a diverse set of complex interactions.
Test
Spot
Optic Flow Stimulus
We interpolated between dual
stimulus data sets to create
versions that represent effects at
intermediate Hot Spot directions.
Optic Flow Responses Predicted By Model With Its Hot Spot Direction
Local Motion Stimuli
SINGLE SITE STIMULI
4 directions
Excitatory
Dual Local Motion Stimuli Reveal Direction Selective Interactions
spks/s
Discharge Rate (spk/sec)
Optic Flow Responses
We then recorded the responses of these
neurons to 30o X 30o patches of local
planar motion by presenting dot pattern
motion in four cardinal directions on an
otherwise blank screen.
(derived from single site, local motion data)
Center Hot Spot Up
We applied the genetic algorithm
to modify the dual Gaussian,
single stimulus receptive field
model for each Hot Spot direction.
Normalized Firing Rate (spks/s)
Here we examine whether simultaneously
presented patches of local motion reveal
MST neuronal response interactions that
might support global pattern selectivity.
We first recorded the responses of MST
neurons in monkeys viewing dot pattern
optic flow stimuli simulating movement in
3D space during centered visual fixation
on a 90o X 90o rear projection screen.
Single site, local motion data yield
dual-Gaussian fits combining an
excitatory and inhibitory mechanism,
or two excitatory, or two inhibitory
mechanisms. In the latter cases, the
two can be so similar as to be
construed as a single mechanism.
2 Site Data Transforms 1 Site Model for Each Hot Spot Direction
Normalized Firing Rate (spks/s)
The optic flow field contains a spectrum of
local directional segments each of which
contains somewhat different directions of
approximately planar local motion.
Dual Gaussian Model
spks/s
The radial pattern of optic flow surrounds
the moving observer and provides robust
cues about the direction of self-movement
as the flow field’s focus of expansion (FOE).
Local Motion Composition Of
The Global Pattern In Optic Flow
Dual Gaussian Model of MST Response Field Can Fit Optic Flow Data
Firing Rate (spks/s)
INTRODUCTION
Single site, local motion stimuli yield
directional response profiles, typically
modeled by combined excitatory and
inhibitory mechanisms.
(
Dual site data is used to modify
the singles data receptive field
model for optic flow having that
local motion direction at that
Hot Spot location.
Dual Site Data
Single to Dual Transform
(Lt-Up Hot Spot Lt)
(Lt-Up w/ Lt; interpolate sites)
)
Single Site Model Interpolated Transform Transformed Model
(Lt-Up w/ Lt + interpolation) (For Flow w/ Lt-Up w/ Lt )
(
)
• Dual simultaneous stimuli reveal
dynamic interactions between sites
throughout the receptive field.
• Fits to optic flow responses can be
improved by transforming models
using dual site response interactions.
• Create a Monte Carlo simulation of dual
site transforms by applying each neuron’s
dual site transforms to a) other sites in the
neuron, & b) all sites in all other neurons.
This work was supported by grants from NEI (R01EY10287, P30EY01319).