In many industrial ML applications, feature engineering consumes the lion’s share of time, energy, and resources. Deep learning promises to replace feature engineering with models that learn end to end from “close-to-reality” data and has convincingly realized this promise for computer vision and natural language processing. But does this promise apply outside of these domains?

Pamela Vagata (Stripe) explains how Stripe has applied deep learning techniques to predict fraud from raw behavioral data. Since fraud detection is a critical business problem for Stripe, the company already had a well-tuned feature-engineered model for comparison. Stripe found that the deep learning model outperforms the feature-engineered model both on predictive performance and in the effort spent on data engineering, model construction, tuning, and maintenance. Join in to discover how common industry practice could shift toward deeper models trained end to end and away from labor-intensive feature engineering.

This session was recorded at the 2019 O'Reilly Artificial Intelligence Conference in New York.