The effects of data augmentation on NN age prediction performance.
<p>Data augmentation introduces random noise to each sample of the training data in each batch of training. This method can help NN algorithms with a large number of parameters generalize better, by preventing the memorization of the samples in the training set. When augmentation levels grow v...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , , , , , , , , , , , , , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>Data augmentation introduces random noise to each sample of the training data in each batch of training. This method can help NN algorithms with a large number of parameters generalize better, by preventing the memorization of the samples in the training set. When augmentation levels grow very large, however, the signal in the data is overwhelmed by the noise that is added in augmentation, and the algorithm can no longer learn. In our data, we found that augmentation can have dramatic effects on algorithm performance in the brain age prediction task. For example, the resnet NN algorithm, which had poor <i>R</i><sup>2</sup> in the augmentation-free condition, reaches parity with the baseline model at relatively high augmentation levels ( standard error of the mean (SEM), red curves). The lstmfcn NN, which also performs poorly with no augmentation, reached even higher <i>R</i><sup>2</sup> than the baseline model with high levels of augmentation ( SEM, pink curves). However, at these higher levels of augmentation, the data requirements of these two models also increases. Algorithms that were similar in their performance to the baseline in the absence of augmentation improve slightly with the introduction of small amounts of augmentation. For example, the highest <i>R</i><sup>2</sup> reached by any model in these experiments is reached by the blstm1 model at a low value of augmentation ( SEM, gray curves). The relatively-simple mlp4 model architecture that does not perform very well in the absence of augmentation, only becomes worse with the introduction of augmentation (blue curves). Further quantification of these trends is laid out in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013323#pcbi.1013323.s003" target="_blank">S3 Fig</a>, <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013323#pcbi.1013323.s004" target="_blank">S4 Fig</a>, <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013323#pcbi.1013323.s005" target="_blank">S5 Fig</a>.</p> <p>(PNG)</p> |
|---|