Data Augmentation

The performance and accuracy of nearly every machine learning model in existence depend heavily on the relevancy, quality, and quantity of training data. Unfortunately, sufficient data isn’t always available, particularly for niche or emerging use cases. Moreover, even if the data exists, collecting it may be beyond the capabilities of a business.

Provided the business has access to at least some training data, it can instead leverage data augmentation.

What is Data Augmentation?

Data augmentation uses existing data from a data set to artificially generate new data points. In order to avoid redundancy and maintain distribution, it will typically also make minor changes to data. If the data is an image set being used for computer vision, for instance, it may resize, crop, recolor, or flip new images.

A business might even take things a step further and use deep learning algorithms for automatic data augmentation, allowing for faster generation of more accurate data and even allowing the interspersal of synthetic data.

Augmented Data vs. Synthetic Data

Augmented data is derived entirely from existing data. While the data may undergo certain modifications for the sake of diversity, that link to the real world still exists. Synthetic data, on the other hand, is completely artificial.

Instead, synthetic data is created algorithmically, typically through the use of deep neural networks and generative adversarial networks. While one might assume this makes synthetic data generally inferior to augmented data, the opposite may actually be true. In recent years, synthetic data technology has grown significantly more advanced, and it now allows users to generate large data sets specifically tailored to their needs.

According to analyst Gartner, synthetic data is gaining such widespread acceptance that by 2024, 60 percent of the data sets used in the development of analytics and artificial intelligence will be synthetic. The use cases for these data sets include, but are not limited to:

  • Training neural networks and deep learning models.
  • Lessening bias in existing data sets.
  • Protecting user privacy.
  • Complying with industry regulations and frameworks.

Why is Data Augmentation Important?

A lack of sufficient training data remains one of the most significant challenges in the machine learning space, and data collection can only go so far in addressing the problem. For some training use cases, large data sets simply do not exist. Alongside synthetic data, data augmentation helps bridge the gap.

Rather than going through the painstaking, expensive process of data selection, collection, and labeling, businesses can instead generate high-quality samples from the data that is already available to them. Data augmentation can also be used to enrich existing data, improving diversity by ensuring a wider selection of data points. This is particularly significant for use cases where data sanitization and pre-processing might reduce the representability of a data set.

To summarize, the benefits of data augmentation in machine learning include:

  • Reduced tendency towards overfitting.
  • Lower data redundancy.
  • Improved overall model accuracy.
  • Prevention of data scarcity.
  • Introducing more training data into existing models.
  • Resolution of class imbalance issues in classification.
  • Reduced the operational burden and cost of collecting, labeling, and cleaning raw data.
  • Enabling rare event prediction and allowing a model to train on ‘edge’ cases.

It’s important to note that, while beneficial, data augmentation also isn’t without its drawbacks. If the original data set contains any inaccuracies or biases, the augmented data will duplicate these flaws. Quality assurance for augmented data can also be surprisingly expensive, and it can be incredibly challenging to find an effective data augmentation system and approach.

How Does Data Augmentation Work?

There exists a multitude of data augmentation techniques for each type of data that might be found in a data set. There are also several advanced techniques that may be broadly applied regardless of format.

Data Augmentation for Image Classification

For images and video, data augmentation is frequently combined with synthetic data created by generative adversarial networks. Some of the transformation and alteration techniques used in this process include:

  • Padding.
  • Re-scaling.
  • Rotation
  • Horizontal and vertical flipping.
  • Translation along the X and/or Y axis.
  • Cropping.
  • Modification of RGB color channels.
  • Modification of brightness, contrast, and shading.
  • Addition of visual noise.
  • Erasure.
  • Sharpening and blurring.
  • Mixing and blending multiple images.

Data Augmentation for Natural Language Processing

While data augmentation is quite popular in computer vision, its use is much less widespread for natural language processing. This is likely due as much to the relative abundance of text-based training data as it is to the challenges of augmenting a data set built atop a complex language. Even so, there are still several augmentation techniques one might use here:

  • Changing the position of words or sentences.
  • Replacing words with synonyms.
  • Paraphrasing sentences using the same or similar-meaning words.
  • Inserting new words at random.
  • Deleting words at random.
  • Re-translating text from one language to another.
  • Embedding new words based on context.

Data Augmentation for Audio

Similar to NLP, the use cases for audio augmentation are relatively sparse, as are the available augmentation techniques:

  • Injection of random or Gaussian noise into the data set as a whole.
  • Shifting audio left or right at random intervals.
  • Increasing or reducing the speed of the audio.
  • Increasing or reducing the pitch of the audio.

Advanced Data Augmentation Techniques

Advanced models and techniques for data augmentation include:

  • Adversarial training, which generates data points intended to disrupt a machine learning model, then injects them into a data set for training purposes.
  • Generative Adversarial Networks (GANs), which are capable of analyzing existing data sets and automatically generating new examples that resemble existing data.
  • Neural-style transfer models, which can deconstruct and reconstruct images while also separating style and context.
  • Reinforcement learning models, which allow software agents to independently learn, explore, and make decisions within a sandboxed environment.

Potential Applications of Data Augmentation

Given the growing prominence of artificial intelligence and machine learning, it should come as no surprise that data augmentation is also experiencing a surge in growth. After all, it can prove invaluable for virtually any industry. With that said, data augmentation is especially useful in the following sectors:

  • Healthcare: Rather than having to go through the process of acquiring and labeling sensitive medical images, physicians can instead use data augmentation to create diverse data sets from a relatively limited base.
  • Autonomous Vehicles: Given that it’s infeasible to road test a self-driving car in an urban or residential environment, developers instead rely on sophisticated simulations and scenarios, leveraging a combination of deep learning, data augmentation, and synthetic data.
  • Speech Recognition: For all that technology has advanced in the past several years, speech recognition algorithms still leave much to be desired. Data augmentation can help improve performance and accuracy of these models, allowing them to operate with much greater efficacy and efficiency.
  • Natural Language Processing: Although text data augmentation is still relatively underexplored compared to other use cases, there are still scenarios where it can prove invaluable – particularly when improved model performance is a priority.
  • Computer Vision: The broader field of computer vision may well represent the most significant use case for data augmentation – and with good reason. Data augmentation has been shown on multiple occasions to both enhance accuracy and reduce overfitting in computer vision. As mentioned earlier, it can also be leveraged to ensure that a computer vision model receives sufficient training for edge cases.