Synthetic Data Generation

What is synthetic data?

Synthetic data generation is the practice of creating a model to generate data that reflects a real world system. Synthetic data can be a cheap and practical way to generate large datasets, and simulations are one of the best ways to generate synthetic data.

By creating a simulation that models the real world, the results of the simulation can be used in place of, or in addition to, directly observed data. In this way a small investment of 'data' can be supercharged with a simulation to create a much larger dataset, and a more performant machine learning (ML) system.

Using synthetic data

Synthetic data is used in numerous applications:

• →

  Training ML Models: Most modern ML applications are data hungry, requiring large datasets for training and prediction. It's not always feasible to use real data, either because it's too expensive, the data is too noisy, or you're building something new for which reliable data sets don't exist. For example synthetic data is frequently used to train self-driving cars, where realistic simulations create synthetic environments to train the models.

• →

  Privacy: Real world datasets can contain identifying information that is hard if not impossible to anonymize. Generating synthetic data from a model removes the risk of compromising information, since none of the data will come from real people.

• →

  Counterfactuals: Given a real world system, you might want to generate data about alternatives to the real world system (counterfactuals) to train predictive models for contingencies. For instance, when monitoring cloud infrastructure, what would happen if a server goes down - what would the data signal patterns look like? Actually bringing down a server and recording the data might be prohibitively expensive, but creating a digital twin of your cloud infrastructure and simulating outages can create robust datasets.

Generating synthetic data

Simulation models in HASH can be run one or many times and their outputs used as synthetic data. Models which incorporate stochasticity or use ranges of realistic input parameters can help produce robust, realistic synthetic data which is representative of the real-world.