What Is Synthetic Data?

Synthetic data is any information manufactured artificially which does not represent events or objects in the real world. Algorithms create synthetic data used in model datasets for testing or training purposes. The synthetic data can mimic operational or production data and help train machine learning (ML) models or test out mathematical models.

Synthetic data offers several important benefits: it minimizes the constraints associated with the use of regulated or sensitive data, it can be used to customize data to match conditions that real data does not allow, and it can be used to generate large training datasets without requiring manual labeling of data.

Why Is Synthetic Data So Important?

Developers often require large, accurately labeled datasets when training AI models. Neural networks achieve higher accuracy when trained with more diverse data. However, it may be unrealistically time-consuming and expensive to collect and label these large datasets with thousands or even millions of objects.

Synthetic data can significantly bring down the cost of generating training data. For example, a training image that costs $5 if sourced from a data labeling service might cost as little as $0.05 if generated artificially.

In addition to lowering costs, synthetic data helps address privacy issues associated with potentially sensitive data sourced from the real world. It can also reduce bias compared to real data, which might not accurately reflect the full range of information about the real world. Synthetic data can provide greater diversity by including rare cases that represent realistic possibilities but may be difficult to source from authentic data.

Advantages of Synthetic Data

Data scientists shouldn’t mind if the data they use is authentic or synthetic, as long as it represents accurate patterns, is balanced, unbiased, and high quality. Synthetic data allows for enrichment and optimization, which allows data scientists to unlock several advantages:

  • Data quality – in addition to being complicated and expensive to collect, real-world data is often full of errors, containing inaccuracies or representing a bias that may affect the quality of a neural network. Synthetic data ensures higher data quality, balance, and variety. Artificially-generated data can automatically fill in missing values and apply labels, enabling more accurate predictions.
  • Scalability – machine learning requires massive amounts of data. It is often difficult to obtain relevant data on the necessary scale for training and testing a predictive model. Synthetic data helps fill in the gaps, supplementing real-world data to achieve a larger scale of inputs.
  • Ease of use – synthetic data is often simpler to generate and use. When collecting real-world data, it is often necessary to ensure privacy, filter out errors, or convert data from disparate formats. Synthetic data eliminates inaccuracies and duplicates and ensures all data has a uniform format and labeling.

What Is a Synthetic Dataset?

A synthetic dataset is a dataset containing computer-generated data rather than real-word records. A major use for synthetic datasets is to provide robust, versatile data sufficient for ML training purposes.

Synthetic data must have specific properties to be useful for machine learning models like classification algorithms. A synthetic dataset may contain binary, numerical, categorical data, or unstructured data like images and video. It is important to control the random processes that generate data based on statistical distributions or generative models, to ensure that the result is sufficiently diverse, yet still realistic.

Synthetic data should be customizable. For example, it may be useful to adjust the degree of class separation to simulate harder or easier classification problems. On the other hand, other tasks like regression may benefit from generating data using non-linear processes.

What Are the Use Cases for Synthetic Data?

Synthetic data should accurately represent the original data it augments. High-quality synthetic data can serve as a substitute for actual sensitive production data in a non-production environment (i.e., training, testing, analytics, development, etc.).

Synthetic data also helps data scientists comply with data privacy regulations such as HIPAA, GDPR, CCPA, and CPA. Synthetic data is ideal for safely using sensitive datasets for training or testing purposes. Businesses can extract insights from such data without impacting privacy compliance.

Typical use cases for synthetic data include:

  1. Testing —synthetic test data is easier to generate than rule-based test data and provides flexibility, scalability, and realism. This data is essential for data-driven testing and software development.
  2. AI/ML model training —AI model training increasingly relies on synthetic data. Data synthesis can augment real data and upsample rarer events or patterns, enabling the algorithm to train more effectively. Synthetic training data typically performs better than real-world data and is crucial for building high-quality AI models.
  3. Governance – synthetic data helps remove biases present in real-world data. Synthetic data is also useful for stress-testing an AI model with data points that rarely occur in the real world. Synthetic data is essential for explainable AI and provides insights into how models behave.

How Can Synthetic Data Help Computer Vision?

Here are some ways synthetic data can be useful for computer vision.

Faster, Cost Effective Dataset Creation at Scale

It can be time-consuming and prohibitively expensive to collect the desired visual data from the real world, while ensuring sufficient diversity. Correctly labeling the data points is essential because improperly labeled data might generate an inaccurate outcome. Data collection and labeling or annotation processes could take months, consuming extensive business resources.

Programmatically generated synthetic data doesn’t necessitate manual data collection or annotation efforts. Data scientists can set up synthetic data that includes the proper labels and annotations at the beginning.

Prediction of Edge Cases

Real-world datasets often contain an imbalance, because edge cases, which do not happen frequently in real life, are not sufficiently represented. This can create biases in a machine learning model, because the model might not have enough data to learn how to correctly process the edge case. Edge cases are not necessarily unimportant—they might not occur very often in the real world, but they might be very significant for the process being modeled. 

If a computer vision system learns only from real events, it might not lack sufficient exposure to examples of edge cases. For example, it might not be realistic to naturally expose self-driving cars to the full range of possible events on the road, so the algorithm might not anticipate unusual scenarios that could lead to a car accident. 

Synthetic inputs, such as videos or images of car accidents, can include diverse conditions and events (i.e., light and weather conditions, types and number of vehicles, environments). Autonomous vehicle algorithms trained with diverse synthetic data can produce safer computer vision for cars, accounting for a larger variety of rare real world events.

Data Privacy Protection

Data privacy laws and regulations can affect the collection and storage of visual data. Non-compliance with regulations like GDPR may result in serious legal, financial, or business consequences. Datasets containing private data present a risk because using them in training models could result in a compliance violation. 

For example, users with API access to a facial recognition system can extract images of recognizable faces and match them to the names of real people, representing a potential privacy violation.

Synthetic data avoids the risk of exposing private information because it does not contain any real information traceable to real individuals. Even if a synthetic dataset is based on real data (for example, images of real people), it can preserve the relevant characteristics of the original data without using any identifiable information, eliminating the compliance risk.

Types of Synthetic Data

Data scientists use randomly generated synthetic data to mask confidential information while retaining the statistically relevant characteristics of the original data. Synthetic data generally falls into three main categories:

  • Fully synthetic—retains nothing from the original data. The data generating program typically identifies real-world data characteristics, such as feature density, to estimate realistic parameters. It then randomly generates data based on estimated feature densities, or using generative methods. There is no use of real data with this technique, so it provides robust privacy protection at the expense of data truthfulness. 
  • Partially synthetic—replaces the values of certain selected features with synthetic values while retaining some of the real data, or permutates existing unstructured data. It is also useful for filling in gaps in the original data. Data scientists use model-based and imputation methods to generate partially synthetic data. For structured data subject to privacy restrictions, this technique usually only masks the high-risk or privacy-protected data features. 
  • Hybrid—combines real and synthetic data. Hybrid synthetic data pairs random records from a real dataset with close synthetic records. It offers the benefits of fully and partially synthetic data, providing high utility and privacy protection. The drawback of this data type is the longer processing time and memory required.

Synthetic Data Challenges

Synthetic data offers compelling benefits, but it is not easy to realize them. Synthetic data generation requires highly-skilled AI specialists who understand how data works and can use the sophisticated tools that generate and analyze datasets. Organizations must also establish a framework to validate the success of their data generation projects.

Here are some of the challenges involved in generating synthetic data:

  • Realism—synthetic data must accurately reflect the original, real-world data. However, business departments, customers, or auditors may also require assurances of privacy preservation. It can be difficult to generate realistic data that doesn’t expose actual private data. On the other hand, if the synthetic data is not sufficiently accurate, it won’t reflect the patterns crucial to the training or testing project. Modeling efforts based on unrealistic data cannot generate useful insights. 
  • Bias—often creeps into ML models trained on artificially generated datasets. Both real-world and synthetic data may contain an inherent or historical bias. If the synthetic data accurately mimics the original, it can reproduce the same biases in the newly generated data. Data scientists must adjust the ML models to account for bias and ensure the synthetic dataset is more representative.
  • Privacy—some types of synthetic data are based on real world data. If synthetic data generated by the model is too similar to the original data, this could create privacy issues. This is especially true if the original data used contains personally identifiable information (PII), which may be subject to privacy protection regulations.

Methods for Generating Synthetic Data

To generate synthetic data, data scientists need to create a robust model that models a real dataset. Based on the probabilities that certain data points occur in the real dataset, they can generate realistic synthetic data points. 

Neural networks are especially adept at learning an underlying data distribution and generalizing it. This enables a neural network architecture to create data points that are similar, but not identical, to samples from the original distribution. Here are a few state of the art neural techniques used to generate synthetic data.

Variational Autoencoders (VAE)

VAEs are unsupervised generative models that can learn the underlying distribution of data and generate a complex model. They operate by taking an original distribution, transforming it into a latent distribution, and back into the original space (this is known as encoded-decoded). This process results in a “reconstruction error”, which the model aims to minimize. 

VAEs are very useful for continuous data but less effective at categorical data. They are also limited in their ability to generate images or other types of unstructured data.

Generative Adversarial Networks (GAN)

GAN is a supervised generative model that can be used to generate realistic, highly detailed representations. It works by training two neural networks, one generating fake data points (a generator), and the other aiming to distinguish fake and real data points (a discriminator). Over thousands of training cycles, the generator becomes more and more successful at generating highly realistic fake data points that can “fool” the generator. 

GANs are especially successful at synthetic generation of images, videos, and other unstructured data. Their weakness is that they require specialized expertise to construct and train, and that the model can “collapse” and start to produce a limited set of very similar fake data points.

Neural Radiance Field (NeRF)

NeRF is a method of generating new views from a partially-known 3D scene. The algorithm takes in a set of images, interpolates them, and adds novel viewpoints of the same object. It works by treating the static scene as a continuous 5-dimensional function, and using a fully connected neural network to predict the content of each voxel. For each ray, it provides a predicted volume for one voxel, and so fills in an entire missing picture in the scene. 

NeRF is a very useful way to generate additional, realistic images from an existing image set. Its weaknesses are that it is slow to train, slow to render, and might generate low-quality or aliased images. Several neural rendering algorithms are now available that address these challenges.

Beyond Synthetic Data: Simulated Image Data with Datagen

Synthetic data encompasses a range of data types—anything not collected manually. Techniques for creating synthetic data range from the simple (i.e., pasting subjects onto various backgrounds to produce variation) to advanced AI methods like the ones covered above. 

Most data creation techniques are static, which means data scientists have to regularly adjust and regenerate them for new training objectives, distributions, or parameters. Synthetic visual data, in particular, can be intensive to modify. The standard methods typically focus on a single task or scenario and don’t support real-time, responsive machine learning.

What is Simulated Data?

Simulated data is a sophisticated form of synthetic data that uses a virtual camera to generate photorealistic simulations that follow the laws of physics. Simulated data includes all the necessary annotations and dimensions, producing realistic 3D data.

Datagen, a leader in synthetic data generation, has established a massive set of 3D asset libraries with real-world human-centered scans, ensuring realism and robust variation in simulated data. The hyper-real simulations are uniquely flexible and customizable. Datagen can virtually capture photographic information of an environment in different conditions using various angles and lenses, providing insights into the behavior of 3D assets. 

Unique Features of Simulated Data

Compared to static synthetic data, simulated data allows organizations to:

  • Place different furniture in the simulated space
  • Adjust lighting conditions
  • Modify layouts, colors, and textures
  • Capture infrared images and depth maps
  • Capture edge cases that rarely occur in the real world

All simulated data is based on high-quality scans of real 3D objects, representing plausible, realistic scenarios. Synthetic and simulated data help address privacy concerns, iron out inherent biases, and dramatically reduce the cost of gathering data at scale compared to manual techniques. 

The Datagen solution is a fully customizable sandbox for exposing systems to dynamic environments of 3D spaces, people, and objects.