Does the term "data augmentation" imply increasing the training dataset?

Question

I have a manuscript that has been reviewed and one of the reviewers commented on my use of the term " data augmentation", saying that it might not be the appropriate term in my case (explained below).

I collected a large dataset of short audio files which are used to train a Convolutional Neural Network. Before being used as model input, each audio file is processed through a pipeline that mixes it with other audio files, changes some of the sound properties (SNR ratio, distorting the audio ...) and finally converts it into a mel-spectrogram. I only modify the existing file and I do not increase my training dataset but I refer to this processing as "data augmentation".

I did not find any definitive definition of what is data augmentation. For instance, Salamon and Bello, 2016 define data augmentation as

the application of one or more deformations to a collection of annotated training samples which result in new, additional training data

However, it appears to me that the increase in the training dataset is only a byproduct of the data augmentation.

In any case, I would really appreciate it if you could confirm or not my use of "data augmentation" and I would be grateful if you can provide a reference for this.

Answer 1

I'm not familiar with any "authoritative" single definition somewhere, or not sure who used the term first, but I would personally indeed agree with the reviewer you mention. In fact I've made similar comments myself as a reviewer of a paper once (but it was in an entirely different field, nothing related to audio at all, so I probably wasn't the reviewer in question). Personally, when I think of data augmentation, the first example that jumps to mind is adding rotated/reflected versions of images to the training data.

A quick google search suggests that, for example, Shorten and Khoshgoftaar (2019) similarly associate data augmentation with increases in training data in their survey paper on data augmentation for images.

What you describe sounds more like "data preprocessing" rather than data augmentation to me, but I'm not 100% sure exactly what the purpose of your modifications is (which I suppose might change my opinion). Data preprocessing can be almost any procedure that is applied to your original data, like... feature selection, converting images to grayscale if you expect colour to be irrelevant to your task, downscaling images if they're too big and you don't mind decreasing the resolution, etc.