Data augmentation role
Data is paramount in any AI project. This is especially true when using the artificial neural network (ANN) algorithm, also known as DL. The success or failure of a DL project is primarily due to the input data quality.
One primary reason for the significance of data augmentation is that it is relatively too easy to develop an AI for prediction and forecasting, and those models require robust data input. With the remarkable advancement in developing, training, and deploying a DL project, such as using the FastAI framework, you can create a world-class DL model in a handful of Python code lines. Thus, expanding the dataset is an effective option to improve the DL model’s accuracy over your competitor.
The traditional method of acquiring additional data is difficult, expensive, and impractical. Sometimes, the only available option is to use data augmentation techniques to extend the dataset.
Data augmentation methods can increase the data’s size tenfold. For example, it is relatively challenging to acquire additional skin cancer images. Thus, using a random combination of image transformations, such as vertical flip, horizontal flip, rotating, and skewing, is a practical technique that can expand the skin cancer photo data.
Without data augmentation, sourcing new skin cancer photos and labeling them is expensive and time-consuming. The International Skin Imaging Collaboration (ISIC) is the authoritative data source for skin diseases, where a team of dermatologists verified and classified the images. ISIC made the datasets available to the public to download for free. If you can’t find a particular dataset from ISIC, it is difficult to find other means, as accessing hospital or university labs to acquire skin disease images is laced with legal and logistic blockers. After obtaining the photos, hiring a team of dermatologists to classify the pictures to correct diseases would be costly.
Another example of the impracticality of attaining additional images instead of augmentation is when you download photos from social media or online search engines. Social media is a rich source of image, text, audio, and video data. Search engines, such as Google or Bing, make it relatively easy to download additional data for a project, but copyrights and legal usage are a quagmire. Most images, texts, audio, and videos on social media, such as YouTube, Facebook, TikTok, and Twitter, are not clearly labeled as copyrights or public domain material.
Furthermore, social media promotes popular content, not unfavorable or obscure material. For example, let’s say you want to add more images of parrots to your parrot classification AI system. Online searches will return a lot of blue-and-yellow macaws, red-and-green macaws, or sulfur-crested cockatoos, but not as many Galah, Kea, or the mythical Norwegian-blue parrot – a fake parrot from the Monty Python comedy skit.
Insufficient data for AI training is exacerbated for text, audio, and tabular data types. Generally, obtaining additional text, audio, and tabular data is expensive and time-consuming. There are strong copyright laws protecting text data. Audio files are less common online, and tabular data is primarily from private company databases.
The following section will define the four commonly used data types.