In this project, I train convolutional neural networks (CNNs) on synthetic data generated by a variational autoencoder (VAE) for the task of handwritten digit recognition. I look at how different mixtures of real and synthetic training data affect the performance of these CNNs. Ultimately, I found that adding synthetic data generally did lead to a performance improvement, though this improvement was not as pronounced as just training on redundant data.
The full write up is the file, Synthetic_Data_Project_Write_Up.pdf.