Invariant EEG Representation Learning via Adversarial Inference

This is the invariant EEG representation learning convolutional neural network (CNN) implementation, which uses adversarial training to monitor and/or censor nuisance-specific leakage in learned representations. The aim is to perform adversarial censoring during the CNN training procedure such that the network can learn nuisance-invariant representations within the discriminative CNN setting. Implementation is in Python using Keras with Tensorflow backend.

Usage

An example execution is as follows:

from AdversarialCNN import AdversarialCNN

net = AdversarialCNN(chans = ..., samples = ..., n_output = ..., n_nuisance = ..., architecture = ..., adversarial = ..., lam = ...)

net.train(train_set, validation_set, log = ..., epochs = ..., batch_size = ...)

Boolean parameter adversarial = True trains the network via adversarial censoring. If False, then an adjacent adversary network is simply trained to monitor nuisance-specific leakage in the representations. Parameter lam indicates the adversarial regularization weight embedded in the loss function. Parameter architecture defines the regular CNN blocks excluding the final dense layer (i.e., linear classification layer), which can be modified arbitrarily. Default implementations include architecture = 'EEGNet' from (Lawhern et al. 2018), as well as architecture = 'DeepConvNet' and architecture = 'ShallowConvNet' from (Schirrmeister et al. 2017), which are well-known CNN architectures used for EEG classification.

To use the train function, both train_set and validation_set should be three-element tuples (i.e., x_train, y_train, s_train = train_set). Here, the first element x_train is the EEG data of size (num_observations, num_channels, num_timesamples, 1), y_train are the one-hot encoded class labels (e.g., for binary labels will have a size (num_observations, 2)), and s_train are the one-hot encoded nuisance labels (e.g., for 10-class nuisance labels will have size (num_observations, 10)). Variable log indicates the directory string to save the log files during training.

Paper Citation

If you use this code in your research and find it helpful, please cite the following paper:

Ozan Ozdenizci, Ye Wang, Toshiaki Koike-Akino, Deniz Erdogmus. "Learning Invariant Representations from EEG via Adversarial Inference". IEEE Access, 2020. https://dx.doi.org/10.1109/ACCESS.2020.2971600

Acknowledgments

Ozan Ozdenizci and Deniz Erdogmus are partially supported by NSF (IIS-1149570, CNS-1544895, IIS-1715858), DHHS (90RE5017-02-01), and NIH (R01DC009834).