DAPE (Domain Aligned Private Encoders) is a Framework for learning from multiple data-sources. Thereby, multiple encoders are jointly trained on data of different data-sources while using one common classifier. An MMD Loss is used for aligning the learnt representations of the encoders during train-time.
If you find this work helpfull, please cite the paper, this repository is based on
@INPROCEEDINGS{DAPE2022Bethge,
author={Bethge, David and Hallgarten, Philipp and Grosse-Puppendahl, Tobias and Kari, Mohamed and Mikut, Ralf and Schmidt, Albrecht and Özdenizci, Ozan},
booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={Domain-Invariant Representation Learning from EEG with Private Encoders},
year={2022},
volume={},
number={},
pages={1236-1240},
doi={10.1109/ICASSP43922.2022.9747398}}
The DAPE Framework consists of mutliple encoders, one for each-data source. The encoder backbone can be chosen freely, it is also possible to use different encoders for different data-sources. The only constrain consists in the encoders outputs needing to have the same shape. The latent representations output by the encoders are used as input for a shared classifier, which makes predictions for each sample of each data-source.
In order to train the Framework, you can use the methods in pipeline_funcs.py, i.e. train and test to build yourself a custom pipeline. Else you can get a head start by using the method pipeline in the file pipeline.py to train and test a DAPE Framework. For more information on the methods parameters, please refer to the comments in the files.
If you have any question or ideas feel free to reach out on me!