Data augmentation is a technique widely used in machine learning and computer vision to increase the size of training datasets by creating new data points from the existing ones. These new data points are variations of the original data, and they help improve the robustness and generalization of machine learning models. While data augmentation doesn't exactly create something from nothing, it generates diverse examples from the available data, enhancing the model's ability to learn patterns and features.

Here are some substantial and reliable advancements in data augmentation techniques up to my knowledge cutoff date in September 2021:

Here an answer without ChatGPT ;)

Data augmentation does not create something from nothing, you still have a basic dataset, which you use as input to create similar data. By similar data, I mean data that fulfills the distribution (mean, variance, etc.) of the basic dataset.

In Machine learning, it is definitely NOT something that I would recommend as it enhances the overfitting problem.

If you want to have some predictions for digital twins/virtual patients, data augmentation might be an interesting approach, as you may observe evolutions of "other", artificially created data that gives you an idea what would happen if this "artificial data" were to be measured one day in reality.

Another facet of data augmentation in general might be to relax a bit the contraints on fulfilling the same distribution, allowing for augmented data beyond the inititial basic dataset. But here, both for machine learning and other contexts, you have to pay attention!

There have been several substantial and reliable advancements in data augmentation techniques. Some of them include:

These advancements have proven to be reliable and effective in improving model training, generalization, and robustness, leading to better performance in various domains and tasks.