Could synthetic data be the solution to rapidly train artificial intelligence (AI) algorithms? There are advantages and disadvantages to synthetic data; however, many technology experts believe that synthetic data is the key to democratising machine learning and to accelerate testing and adoption of artificial intelligence algorithms into our daily lives. 

What is synthetic data?

When a computer artificially manufactures data rather than measures and collects it from real-world situations it’s called synthetic data. The data is anonymized and created based on the user-specified parameters so that it’s as close as possible to the properties of data from real-world scenarios.

One way to create synthetic data is to use real-world data but strip the identifying aspects such as names, emails, social security numbers and addresses from the data set so that it is anonymized. A generative model, one that can learn from real data, can also create a data set that closely resembles the properties of authentic data. As technology gets better, the gap between synthetic data and real data diminishes.

Synthetic data is useful in many situations. Similar to how a research scientist might use synthetic material to complete experiments at low risk, data scientists can leverage synthetic data to minimise time, cost and risk. In some cases, there isn’t a large enough data set available to train a machine learning algorithm effectively for every possible scenario so creating a data set can ensure comprehensive training. In other cases, real-world data cannot be used for testing, training or quality-assurance purposes due to privacy concerns, because the data is sensitive or it is for a highly regulated industry.

Advantages of synthetic data

Huge data sets are what powers deep learning machines and artificial intelligence algorithms that are expected to help solve very challenging issues. Companies such as Google, Facebook and Amazon have had a competitive advantage due to the amount of data they create daily as part of their business. Synthetic data allows organisations of every size and resource levels the possibility to also capitalise on learning that is powered by deep data sets which ultimately can democratise machine learning. 

Creating synthetic data is more efficient and cost-effective than collecting real-world data in many cases. It can also be created on demand based on specifications rather than needing to wait to collect data once it occurs in reality. Synthetic data can also complement real-world data so that testing can occur for every imaginable variable even there isn’t a good example in the real data set. This allows organisations to accelerate the testing of system performance and training of new systems.

The limitations for using real data for learning and testing are reduced when using fabricated data sets. Recent research suggests that it is possible to get the same results using synthetic data as you would with authentic data sets.

Disadvantages of synthetic data

It can be challenging to create high-quality synthetic data especially if the system is complex. It’s important that the generative model creating the synthetic data is excellent or the data it generates will be affected. If synthetic data isn’t nearly identical to a real-world data set, it can compromise the quality of decision-making that is being done based on the data.

Even if synthetic data is really good, it is still a replica of specific properties of a real data set. A model looks for trends to replicate, so some of the random behaviours might be missed.     

Applications of synthetic data

Whenever privacy concerns are an issue such as in the financial and healthcare industries or an enormous data set is required to train machine learning algorithms, synthetic data sets can propel progress. Here are just a few applications of synthetic data:

• Synthetic data with record-level data can be used from healthcare organisations to inform care protocols while protecting patient confidentiality. Simulated X-rays are combined with actual X-rays to train AI algorithms to identify conditions.
• Fraudulent activity detection systems can be tested and trained without exposing personal financial records.
• DevOps teams use synthetic data to test software and ensure quality.
• Machine learning algorithms are often trained with synthetic data.
• Waymo tested its autonomous vehicles by driving 8 million miles on real roads plus another 5 billion on simulated roadways. Other automakers are using video games such as Grand Theft Auto to aid its self-driving technology.

While synthetic data isn’t fool proof, it is an important tool to augment machine learning algorithms when real data is too expensive to collect, inaccessible due to privacy concerns or incomplete.