ERPLAB Studio Panels: Multivariate Pattern Classification
The Multivariate Pattern Classification panel allows you to perform pattern classification (decoding) on the selected BESTsets. See the Decoding Tutorial for details of how ERP decoding works in general, as well as ERPLAB's implementation. This page provides a simple description of the options in the panel.
Each bin is considered a "class" to be distinguished from the other selected bins/classes. You can decode across All bins in the BESTset, or you can select a Custom subset by listing the bins in the Class ID text box. In this example shown above, the decoding will attempt to distinguish among bins 5, 6, 7, and 8. The Chance level is automatically calculated as 1 divided by the number of classes.
You can select which channels will be used for decoding with the Channels text box.
The user must specific the number of Cross-validation Blocks, which is the same thing as the number of averages that will be created for each class. The number of trials per average will then be determined by assigning the available trials equally among the number of averages (except for the constraints described next).
It is ordinarily important for the number of trials to be the same in each average for all the bins/classes being decoded. This is achieved by checking the box labeled Equalize Trials, which “floors” the number of trials. The option for flooring Across Classescauses the number of trials per average to be the same for all classes (using the largest number possible while maintaining the same number across classes). By default, the flooring is done separately for each BESTset. However, if Across BESTsets is selected the smallest number of trials across all selected bins and BESTsets will be used. This latter option is typically used when the user needs to ensures that the noise level is controlled across participants (e.g., when different groups are being compared).
Alternatively, the user can select Manual Floor and enter the number of trials per average that should be used (which may not be greater than the number available). This option is used when the appropriate floor cannot be automatically determined by the software.
The number of trials available in each bin and the number of trials that will be used per average can be seen in the table near the bottom of the panel. This is shown separately for each BESTset.
The user can also specifiy the number of iterations. A larger number will typically yield a more precise estimate of decoding accuracy.
Several options can be controlled by clicking the Options button, which brings up the window shown below.
One option is the decoding method. Currently, we offer support vector machine (SVM) decoding and the cross-validated Mahalanobis distance.
For the SVM option, multiclass decoding can be performed using either one-versus-one decoding (in which a separate decoder is trained for each pair of classes) and one-versus-all decoding (in which a separate decoder is trained for each class relative to the combination of the other classes).
The Decoding Time Range option is used to select which time points are decoded. Note that the decoding is performed completely separately at each time point.
The Subsample option is used to make the decoding finish faster by decoding every Nth time point.
The speed of decoding can also be increased by selecting Yes for Parallelization. When this is turned on, Matlab will see how many processing cores are in your CPU and attempt to use most of them. Note that it may take several seconds for Matlab to get the cores set up once you start decoding. This option requires that “Parallel Computing Toolbox” is already installed in MATLAB. This toolbox can be purchased from the makers of Matlab.