Dynamic Adversarial Data Collection


August 11, 2022


Are you looking for better training data for your models? Let me tell you about dynamic adversarial data collection!

I had a large enterprise customer asking me to incorporate this workflow into a Hugging Face private hub demo. Here are some resources I found useful: Chris Emezue put together a blog post: “How to train your model dynamically using adversarial data” and a real-life example using MNIST using Spaces.

If you want an academic paper that details this process, check out: Analyzing Dynamic Adversarial Training Data in the Limit. By using this approach, this paper found models made 26% fewer errors on the expert-curated test set.

And if you prefer a video — check out my Tik Tok: