Biomedical Engineering Reference
In-Depth Information
∼
10%. However, this approach requires several libraries for the
program R (http://cran.r-project.org/) for rapid classification.
Examples are available on our website (spikelab.jbpierce.org/
Resources)
and
have
been
described
in
detail
in
previous
publications
(5)
.
If neurons have a shared predictive relationship, then they
should have shared predictions that are stronger than would be
expected in trial-shuffled data. This is readily quantified by the
following code:
for i = 1:10000
pred1 = zeros(1,100); pred2 = zeros(1,100);
pred1(find(rand(1,100)>0.7))=1; % 70% prediction
pred2(find(rand(1,100)>0.6))=1; % 60% prediction
match(i) = length(find(pred1==pred2));
end
The distribution of the variable match provides a probability
distribution that is used to assess the significance of predictions.
We applied this approach to two neurons from dmPFC
and motor cortex. Although these two neurons provided small
amounts of predictive information (0.005 bits for the motor cor-
tex neuron, 0.03 bits for the dmPFC neuron;
Fig. 7.8B
), we
found that these neurons shared predictions 59% of the time,
more than could be expected by chance (44
0. 05).
These data indicate that neurons that were weakly predictive of
behavior could still share predictive relationships.
Across our population, we compared neuronal predictions
with neuronal predictions from
within class
(i.e., within fast and
with slow RTs) trial-shuffled data (
Fig. 7.8C)
. We found that
improvements in classification of 9% over random data corre-
sponded to
p
±
9%,
p
<
0. 05. We found that 12 (of 127; 10%) predictive
interactions were greater than could be expected by chance. We
would expect to find this number of significant predictive interac-
tions at
p
<
0. 05 by chance (X
2
= 2.15,
p
0. 14).
We also compared predictive information on a trial-by-trial
basis between dmPFC and motor cortex. The population of 10
dmPFC neurons provided 0.2 bits of information, whereas the
population of 11 motor cortex neurons provided 0.5 bits of infor-
mation. When predicting fast RTs, dmPFC and motor cortex
shared predictions (76%) that could be explained by chance (
p
<
<
<
0. 05 at 76%). On the contrary, when predicting slow RTs,
dmPFC and motor cortex shared predictions (86%) were higher
than could be explained by mere correlations with RT (
p
<
0. 05
at 83%).
The use of statistical pattern recognition to explore trial-
by-trial relationships in predictions between neurons should be
approached carefully. This analysis is complex and reliant on
