Maskul_NN

Scripts Description

• Normalizer: This function normalizes the "raw" input data using a MinMax approach. The idea is to make each feature of the input to belong the interval [0,1]; each feature maximum value is marked as 1, while each feature minimum value is marked as 0.

• do_windows: This function is required to create the windows according to the approach one wants to use. It is possible to create either "static" windows (e.g. discarding transitions) or dynamic windows, overlapping windows (sliding windows).

• do_folds_inter_subjects: This function creates the folds to feed the NN by mixing up the whole set of subjects.

• do_folds_intra_subject: This function creates the folds to feeed the NN on the same patient.

Windows Creation

Parameters

• sliding: If true, it allows to use the sliding windows approach.

• sliding_pace: The name says everything.

• predict_central_sample: If true, it allows to predict the central sample between two overlapping windows. Indeed, the pace must be smaller than the window_size.

• discard_transitions: If true, it eliminates the windows in which there is a transition in the label, according to a threshold.

• threshold: It is required to establish which window needs to be discarded. A threshold equal to 1 means that the labels in the window must be all equal, otherwise the window will be discarded.

NN Description

Hyperparameters

• Learning Rate: The net has to find, step-by-step, an optimum weight in the set of weight. The optima could be either global or local, and the net must be able to find them. To do so, the net has to "change" of a certain amount in each searching step, and this quantity is called learning rate. It has been set to a value of 0.1 by default.

• Batch Size: This value defines the number of samples which will be propagated through the network per time. For instance, if the training set has 512 samples, and the batch size is set to 32, then the network will be trained 32 samples per time, until the end of the 512 total samples. This value has been set to 32.

• Number of Folds: A neural network must be trained many times by "shuffling" the data with which has been fed. This is done in order to increase the variability of data, as well as to see "better" how the network works. This approach is mandatory, and a value of 10 folds is suggested.

• Validation Split: It is the percentage of the training set allocated for creating the dev set (development set). We decided to split the initial training set with a 80/20 ratio, thus its value has been set to 0.2.

• Test Validation Split: It is the percentage of the training set allocated for creating the test val set (test validation set). Actually, this value is not used if we are using three files from the folding process (Train, Test Learned, and Test Unlearned). See the Folding Procedure for a better explanation.

• Samples per Window (spw): It represents the number of samples composing each window (i.e. each row of the input files) for which a label has been assigned according to the do_windows criteria. It has been set to 20 by default, and it is dependent on how the windowing procedure has been chosen.

• Exclude Features (exclude_features): If true, it allows to exclude the some of the features, selecting them into the variable feature select.

• Include Features Only (include_only_features): On the contrary, if true, it allows to selectively choose the features to be loaded, selecting them into the variable feature select. N.B. include_only_features and exclude_features cannot be both true!

• Epochs

  • Number of epochs (maxepoch): It is the maximum number of epochs the neural network can be trained. It is set by default to 100
  • Patient (maxpatience): It is a threshold which aim is to stop the training process if the accuracy does not improve for a certain number of consecutive epochs. It is set to 10, and it is suggested to not change this value.

Accuracy

Binary Classification

The accuracy of the model for a binary classification, which in this case is related to the 2-levels baographic signal, has made as follows:

1. Inputs and Labels are loaded;
2. Outputs are computed from the inputs
   • If output[i] >= 0.5, that value is set to 1;
   • If output[i] < 0.5, that value is set to 0.
3. The total number of outputs is computed;
4. The number of correct outputs is computed;
5. The percentage of correct over total outputs is the accuracy.

Multiclass

Not implemented yet.