Using Synthetic Data in Data-Centric Machine Learning

In previous chapters, we discussed various approaches to improving data quality for machine learning purposes through better collection and labeling.

Although human labelers, data ownership, and technical data quality improvement practices are critical to data centricity, there are limits to the kind of labeling and data creation that can be performed by individuals or through empirical observation.

Synthetic data has the potential to fill in these gaps and produce comprehensive training data at a fraction of the cost and time of other approaches.

This chapter provides an introduction to synthetic data generation. We will cover the following main topics:

• What synthetic data is and why it matters for data centricity
• How synthetic data is being used to generate better models
• Common techniques used to generate synthetic data
• The risks and challenges with synthetic data use

Let’s start by defining...

Understanding synthetic data

Synthetic data is artificially created data that, if done right, contains all the characteristics of production data.

The reason it’s called synthetic data is that it doesn’t have a physical existence – that is, it doesn’t come from real-life observations or experiments that we create to gather data that we subsequently use to run analysis or build machine learning models on.

A foundational principle of machine learning is that you need a lot of data, ranging from thousands to billions of observations. The amount you need depends on your model.

As we have outlined many times already, when the required volume of data is difficult to come by, one approach is to improve the signal in your data to make it possible to produce accurate and relevant outputs, even on smaller datasets.

Another option is to create synthetic data to cover the gaps. A major benefit of synthetic data is its scalability. Real training data is collected...

Summary

In this chapter, we provided a primer on synthetic data and its common uses. Synthetic data is a key part of the data-centric toolkit because it gives us yet another avenue to much better input data, especially when collecting new data is not feasible.

By now, you should have a clear understanding of the fundamentals of synthetic data and its potential applications. Synthetic data is often used for computer vision, natural language processing, and privacy protection applications. However, the potential of synthetic data goes well beyond these three realms.

Whole books have been dedicated to the topic of synthetic data and we recommend that you dive deeper into the subject if you want to become a true expert in synthetic data generation.

In the next chapter, we’ll explore another powerful technique for improving your data without the need for collecting new data: programmatic labeling.

References

1. https://datagen.tech/guides/synthetic-data/synthetic-data, viewed on 12 November 2022
2. https://blogs.gartner.com/andrew_white/2021/07/24/by-2024-60-of-the-data-used-for-the-development-of-ai-and-analytics-projects-will-be-synthetically-generated/
3. https://unity.com/our-company, viewed on 15 November 2022
4. https://venturebeat.com/ai/unitys-danny-lange-explains-why-synthetic-data-is-better-than-the-real-thing-at-transform-2021-2/, viewed on 15 November 2022
5. Alcorn, M A et al 2019, Strike (with) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects, viewed 13 November 2022: https://arxiv.org/pdf/1811.11553.pdf
6. https://www.tesla.com/VehicleSafetyReport, viewed 13 November 2022
7. Karras T, Aila T, Laine S, Lethtinen J, 2017, Progressive Growing of GANs for Improved Quality, Stability, and Variation: https://arxiv.org/abs/1710.10196
8. Karras T, Aila T, Laine S 2018, A Style-Based Generator Architecture for Generative Adversarial...