Ensemble - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki

The paper 'Interpreting chest X-rays via CNNs that exploit hierarchical disease dependencies and uncertainty labels' as well as the CheXpert dataset paper use ensembles to achieve the best performance on the CheXpert dataset. We employ three different types of ensembles for our project:

For this ensemble we train the following architectures for 5 epochs:
DenseNet121, InceptionResNetV2, Xception, DenseNet169
We use those models to get predictions on the same test generator and simply average the probabilities to get the final prediction.

Similar to the simple average ensemble, the weighted average ensemble also averages the predictions of the ensemble members. The difference is that in this case, the members have individual weights to influence their contribution:

Lastly we try a more sophisticated approach to ensembling: using a meta learner. Here, a learner is trained to optimally combine the predictions of the individual ensemble models: