Biomedical Engineering Reference
In-Depth Information
understanding of procedures such as dimension reduction and
classification. In pattern recognition, one must also worry about
bias
and
variance (32)
, which can readily influence trial-by-trial
predictions.
However, these results indicate that dmPFC and motor cor-
tex populations have functional interactions in their trial-by-trial
predictive information (that is, their predictions about reaction
times) only when predicting slow RTs. This type of analysis sug-
gests that dmPFC neurons and motor cortex neurons functionally
interact on slow but not fast RTs. This novel insight is an exam-
ple of how predictive relationships between populations of neu-
rons can be used to make inferences about how these populations
interact.
10. Network
Interactions:
Synergy and
Redundancy
As an extension of the preceding analyses of shared predictive rela-
tionships, one might ask how the predictive information of a two-
neuron ensemble compared to the predictive information of each
neuron individually. If two neurons provided more information
individually than they do together, then they interact redundantly.
On the other hand, if they provide more information together
than they do individually, then they interact synergistically. This
idea provides a framework
(14, 31, 37)
for interpreting network
interactions (
Fig. 7.9A
)
To assess network interactions, one should:
1. Construct perievent matrices for one neuron.
2. Preprocess the perievent matrices (smoothing and decima-
tion).
3. Reduce the dimensions of the data (via PCA or wavelet-based
methods).
4. Segment the data into a training data and a testing set.
5. Training a classifier from training data to predict the behav-
ioral outcome of testing data.
6. Repeat steps 1-5 for another neuron.
7. Repeat steps 1-5 for the two neurons together.
8. Compare information of the two neurons together with the
information of each neuron individually.
If the predictive information of two neurons were greater
than the sum of the predictive information individually, then we
would call this a
synergistic
interaction. On the other hand, if the
predictive information of two neurons were less than the sum
of the predictive information of each neuron individually, then
we would call this a
redundant
interaction. Finally if the predic-
tive information of two neurons were equal to the sum of the
