Data Augmentation Training by Markus Liedl
Data augmentation is important to fully exploit tagged data. Tagging data manually is expensive and time consuming. Proper data augmentation techniques can make the difference between a useful neural network and a useless one. Yet, whether a technique is proper or not depends on the characteristics of the data and has to be judged or individually designed for each project.
In this one day course we'll look at:
- goals of data augmentation. What should one aim at?
- simple augmentations
- how to measure success
- is data augmentation still an art?
- the principle "imitate the natural noise in the data"
- when should you stop devising augmentations?
- when to not use data augmentation at all.
- advanced augmentations like elastic grid deformations and others.
- many exercises
As data we'll be using images. Many of the openly available datasets can be used for this course: We'll simply take a very small part of the available data and see how good our networks can get.
If you are interested in the training mail me at firstname.lastname@example.org Normally the training will happen in or near Munich, Germany. This training is also available as a workshop.
This course is available in English and German.
Related Deep Learning courses are