NVIDIA Says Kumo Relational Predicts Churn with No Training, Beats Supervised Models
Summary
NVIDIA releases Kumo Relational, a pre-trained model that predicts churn, demand, or fraud directly on relational databases in one forward pass, with no task-specific training or feature engineering. NVIDIA says it beats supervised models on 41 tasks across 15 databases and tops AutoGluon on the SAP SALT benchmark by 8% in mean reciprocal rank.
Key Points
- NVIDIA releases the model code under Apache 2.0 and the weights under the OpenMDW License Agreement, version 1.1.
- NVIDIA says kumo-engine handles databases with 500B+ rows through memory-mapped I/O at up to 5 GB/s.
- Kumo Relational stays stable when up to 75% of foreign-key links are removed, NVIDIA reports.