ACSE (Adversarial Censored Shared Encoder) Framework is a Framework for learning from multiple data-sources. Thereby the concept of adversarial inference is used where a data-source ID d, i.e. the id of the data-source a sample is drawn from, is used as nuisance variable.
The ACSE Framework consists of one encoder, one classifier, and one adversary. The data of each data-source are processed by the encoder and then passed to the classifier and the encoder. The better the adversary can reconstruct the data-source ID from the latent representation, the more the encoder gets penalized. This should lead the encoder to learning domain-invariant representations of the data A visualization of the Framework is shown in the following picture
In Addition in the folder preprocessing
this repository includes notebook for preprocessing the four well known EEG datasets SEED, SEED-IV, DEAP and DREAMER
In order to train the Framework, you can use the method pipeline
in the file model/pipeline.py
to train and test an ACSE Framework. For more information on the parameters you need to pass to the method, please refer to the comments in the file.
If you find this work helpful or want to use it please cite: tbd