Synthetic Data
- [Oslo, Norway]
- Overview
Synthetic data is created by generative AI (GenAI) algorithms, which can be instructed to create bigger, smaller, fairer or richer versions of the original data. Due to how the synthesization process takes place, the data can be augmented to fit certain characteristics.
Synthetic data can be described as fake data, generated by computer systems but based on real data. Enterprises create artificial data to test software under development and at scale, and to train Machine Learning (ML) models.
There are 2 types of data:
- Structured, tabular data
- Unstructured, image and video data
- Pros and Cons of Synthetic Data vs Real Data: Benefits and Challenges
Synthetic data is the perfect fuel for AI and machine learning (ML) development projects. Synthetic data is created by GenAI algorithms, which can be instructed to create larger, smaller, fairer or richer versions of the original data.
Because of the way the synthesis process occurs, the data can be enhanced to fit certain characteristics. In a way, synthetic data is like modeling clay for data scientists and data managers. For example, upsampling minority groups in a dataset can improve the performance of ML models. Alternatively, human biases embedded in the raw data can be removed by introducing fairness constraints into the generation process.
Rather than relying on a training set as a basis, synthetic profiles are independently generated, promoting greater diversity and unpredictability in results. This marks a major change as we move away from relying on data obtained from nature or experiments to using artificially created datasets.
In some cases, these new data sets do not just add to but replace traditional methods, and there are concerns that they may be the result of the ability of AI technology to generate data that may not be entirely accurate or reliable.
While synthetic data is proving invaluable in domains where obtaining real-world data is challenging, expensive, or ethically constrained, concerns are growing about potential threats.
- The Synthesization Process
Artificial intelligence (AI) synthetic data generation is the process of using advanced AI algorithms to create artificial data that has the same statistical characteristics and patterns as existing datasets. This data is generated rather than obtained from direct observations of the real world, which can help ensure the privacy of the original data sources.
When generating synthetic data, it's important to prevent the algorithm from overfitting to the original data, which could cause it to memorize the original data and accidentally leak it during the inference phase.
- Synthetic Data Generation Methods
Synthetic data is used to address data privacy concerns, data availability limitations, and the need for diverse datasets. Research topics include generative adversarial networks (GANs), which have been shown to be successful in generating complex synthetic data like images and text, and artificial intelligence (AI)–based generative models, which use multi-layer neural networks to learn the distribution of real data and generate samples.
- Synthetic Data Quality
There is a lack of research on metrics to evaluate the quality of synthetic data beyond performance metrics like Overall Accuracy (OA) or F1-score.
The ability of AI technology to independently generate data raises questions about the accuracy and reliability of these manually created data sets. We walk a fine line between leveraging the advantages of synthetic data and innovative solutions using GenAI, while ensuring the trustworthiness of the data produced.
As synthetic data sets replace traditional methods of analyzing and collating research data, the need to balance the benefits of diversity and unpredictability with the need for accuracy becomes critical. Addressing these issues associated with the generation of large amounts of synthetic data—is it a particularly valuable necessity in fields where obtaining real-world data is challenging, expensive, or ethically constrained, or is it another threat to research integrity?
- Privacy
Synthetic data research is motivated by privacy concerns, as AI and machine learning algorithms often need personal information to learn. For example, machine learning algorithms often rely on record-level data collected from human subjects, which can lead to privacy concerns and legal risks.
[More to come ...]
