Ryan Drapeau: Battling Fraud with ML at Stripe
The Gradient: Perspectives on AI
00:00
Building a Recurrent Neural Network for Payments
So I joined the company in early 2019 and my first task was to build a recurrent neural network. And I failed at that, of course, otherwise we wouldn't have used xg boost till last summer. We had trouble scaling up these record neural networks beyond to be able to match like the number of events in a sequence that would be necessary to beat xg boost. That scalability problem is when we still face today as we try to explore other architectures like transformers or other sequences. So we pivoted this project from RNNs to then be just a traditional deep neural network model,. kind of like starting a bit easier than trying to go straight to RNN.
Transcript
Play full episode
