The greatest utilization value of synthetic data in 2023

Developing successful AI and ML models requires access to large amounts of high-quality data. However, collecting such data is challenging because:

• Many business problems that AI/ML models can solve require access to sensitive customer data, such as personally identifiable information (PII) or personal health information (PHI). The collection and use of sensitive data raises privacy concerns and leaves businesses vulnerable to data breaches. For this reason, privacy regulations such as GDPR and CCPA limit the collection and use of personal data and impose fines on companies that violate these regulations.
• Some types of data collection are expensive or infrequent. For example, collecting data representing a variety of real-world road events for self-driving cars can be prohibitively expensive. Bank fraud, on the other hand, is an example of a rare event. Collecting enough data to develop machine learning models to predict fraud is challenging because fraud is rare.

As a result, enterprises are turning to a data-centric approach to AI/ML development, including synthetic data, to solve these problems.

Generating synthetic data is inexpensive compared to collecting large datasets and can support AI/deep learning model development or software testing without compromising customer privacy. It is estimated that by 2024, 60% of the data used to develop AI and analytics projects will be generated synthetically.

Figure 1: Synthetic data will exceed real data in the future

What is synthetic data?

Synthetic data, as the name suggests, is data that is artificially created rather than generated by actual events. It is often created with the help of algorithms and is used in a wide range of activities, including as test data for new products and tools, model validation, and AI model training. Synthetic data is a type of data augmentation.

Why is synthetic data important now?

Synthetic data is important because itcan be generated to meet specific needs or conditionsthat are not available in existing (real) data. This is useful in many situations, e.g.

• When privacy requirements limit the availability of data or how the data can be used
• Testing a product for release requires data, but this data either doesn’t exist or is unavailable to testers
• Machine learning algorithms require training data. However, especially in the case of self-driving cars, this data is expensive to generate in real life.

While synthetic data first began to be used in the 1990s, abundant computing power and storage space in the 2010s led to wider use of synthetic data.

What are its applications?

Industries that can benefit from synthetic data:

• cars and robots
• Financial Services
• medical
• manufacturing
• Safety
• social media

Business functions that can benefit from synthetic data include:

• marketing
• machine learning
• Agile development and DevOps
• human Resources

Synthetic data allows us to continue to develop new and innovative products and solutions, otherwise the required data would not exist or be available.

Compare the performance of synthetic data and real data

Data is used in applications, and the most direct measure of data quality is the validity of the data at the time of use. Machine learning is one of the most common data use cases today. MIT scientists wanted to measure whether machine learning models derived from synthetic data could perform as well as models built from real data. In a 2017 study, they divided data scientists into two groups: one used synthetic data and the other used real data. 70% of the time the group using synthetic data was able to produce comparable results to the group using real data. This will give synthetic data advantages over other privacy-enhancing technologies (PETs), such as data masking and anonymization.

Benefits of synthetic data

Being able to generate data that simulates the real thing is a seemingly endless way to create test and development scenarios. While there is a lot of truth to this, it is important to remember thatany synthetic models derived from data can only replicate specific properties of the data, which means they can ultimately only model general trends.

However, synthetic data has several benefits compared to real data:

• Overcome real data usage restrictions: Real data may be subject to usage restrictions due to privacy rules or other regulations. Synthetic data eliminates this problem by replicating all the important statistical properties of real data without exposing it.

These benefits suggest that as our data becomes more complex and more tightly protected, the creation and use of synthetic data will only grow.

Synthetic data generation/creation 101

When determining the best way to create synthetic data, it’s important to first considerthe type of synthetic data you intend to have. There are three broad categories to choose from, each with different advantages and disadvantages:

Fully synthetic: This data does not contain any raw data. This means that it is almost impossible to re-identify any single unit, and all variables remain fully available.

Partial synthesis: Only sensitive data will be replaced with synthetic data. This requires heavy reliance on imputation models. This results in less model dependence, but does mean that some disclosure can be made due to the true values retained in the dataset.

Hybrid Synthesis: Hybrid synthetic data is derived from both real and synthetic data. While ensuring the relationship and integrity between other variables in the data set, the underlying distribution of the original data is studied and the nearest neighbor of each data point is formed. For each record of the real data, an approximate record in the synthetic data is selected, and then the two are concatenated to generate hybrid data.

Two general strategies for constructing synthetic data include:

Drawing numbers from a distribution: This method works by observing a real statistical distribution and replicating fake data. This can also include creating generative models.

Agent-based modeling: To obtain synthetic data in this approach, a model is created to explain the observed behavior and then the same model is used to reproduce random data. It emphasizes understanding the impact of interactions between agents on the entire system.

Deep Learning Models: Variational autoencoders and generative adversarial network (GAN) models are synthetic data generation techniques that increase data utility by feeding the model with more data. Feel free to learn more about how data augmentation and synthetic data support deep learning.

Challenges of synthetic data

While synthetic data offers various benefits that can streamline an organization’s data science projects, it also has limitations:

• Possibly missing outliers: Synthetic data can only mimic real-world data, it is not an exact replica of it. Therefore, the synthetic data may not cover some of the outliers that the original data had. However, outliers in the data can be more important than regular data points, as Nassim Nicholas Taleb explains in depth in his book The Black Swan.

Machine Learning and Synthetic Data: Building AI

Figure 2: The relationship between ML and synthetic data

The role of synthetic data in machine learning is rapidly increasing. This is becausemachine learning algorithms are trained using large amounts of datathat may be difficult to obtain or generate without synthetic data. It can also play an important role in creating algorithms for image recognition and similar tasks that are becoming the baseline for artificial intelligence.

There are several otherbenefits of using synthetic data to helpdevelop machine learning:

UnrealSynth\Windows\UnrealSynth\Content\UserData
    |- images
        |-train
            |- 0.png
            |- 1.png
            |- 2.png
            |- ...
         |-val
            |- 0.png
            |- 1.png
            |- 2.png
            |- ...
    |- labels
        |-train
            |- 0.txt
            |- 1.txt
            |- 2.txt
            |- ...
        |-val
            |- 0.txt
            |- 1.txt
            |- 2.txt
            |- ...
    |- synth.yaml

from ultralytics import YOLO

# Load a model
model = YOLO('yolov8n.yaml') # build a new model from YAML
model = YOLO('yolov8n.pt') # load a pretrained model (recommended for training)
model = YOLO('yolov8n.yaml').load('yolov8n.pt') # build from YAML and transfer weights

# Train the model
results = model.train(data='synth.yaml', epochs=100, imgsz=640)

# Build a new model from YAML and start training from scratch
yolo detect train data=coco128.yaml model=yolov8n.yaml epochs=100 imgsz=640

# Start training from a pretrained *.pt model
yolo detect train data=coco128.yaml model=yolov8n.pt epochs=100 imgsz=640

# Build a new model from YAML, transfer pretrained weights to it and start training
yolo detect train data=coco128.yaml model=yolov8n.yaml pretrained=yolov8n.pt epochs=100 imgsz=640