September 2022 Summaries

In this tutorial, we learn how to use the `gretel-trainer` SDK for conditional data generation in machine learning datasets. This technique allows generating additional labeled examples at a fraction of the cost compared to manual labeling techniques. We demonstrate how to conditionally generate tabular data using Gretel's APIs and AI-based generative models, focusing on the popular MITRE synthetic patient record dataset. The code provided trains a deep learning model on the dataset and samples new synthetic records matching predefined criteria such as race, ethnicity, and gender. This method can be useful for addressing bias in data and correcting class imbalances in various applications like healthcare.

Gretel has launched a Synthetic Data Community Discord server, aiming to provide a platform for individuals interested in synthetic data to share information, communicate, and collaborate. The server will host hackathons, code challenges, and serve as the primary discussion forum for the upcoming synthesize 2023 conference. By choosing Discord, Gretel leverages its voice and video capabilities, support for various platforms, and lack of message retention limit. To join the server, users need to create a Discord account and accept the community code of conduct and Discord rules.

IntroGretel has released an open beta version of its DGAN model, a dedicated time-series model in their API. The DGAN model can generate synthetic time-series data with the convenience of Gretel's API and is suitable for various applications such as smart home sensors or financial data analysis. This release includes a detailed blog post explaining how to use the new DGAN model, including an example demonstrating its basic usage with Python SDK. The DGAN model requires many examples of time series to train effectively, and users can modify the YAML configs as needed. As this is an open beta release, Gretel welcomes feedback from users to improve the experience.

In this case study, Reetam Ganguli, a medical candidate at Brown University and leader of a bioincubator, explains why medical practitioners turn to synthetic data when overcoming challenges with clinical data. Biased data or class imbalance is a significant problem in the medical field due to limited medical data collection from underrepresented demographics, historically low mortality rates for commonly treated diseases, and gender biases stemming from societal and clinical factors. Reetam's team leverages synthetic data to predict postpartum hemorrhages for expecting mothers in Cameroon and Nigeria. Synthetic data can help combat this critical data challenge by generating diverse datasets that enable better research outcomes.

The Gretel Amplify model is a new tool for rapidly generating large volumes of tabular synthetic data using statistical models and hyper-efficient multi-processing implementation. It runs on CPU and can generate data up to 1000x faster than deep learning-based generative models, enabling users to create large numbers of synthetic records very quickly. Amplify is effective at learning and recreating distributions and correlations but typically has a 10-15% drop in synthetic data accuracy compared to Gretel's deep learning models. Some use cases for Amplify include creating synthetic data for load testing applications, mimicking real-world data for pre-production environments, and generating synthetic examples to test machine learning model generalization capabilities.