SignalTech/AI
NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required
NVIDIA Kumo relational model outperforms supervised ML on relational data without training.
Context
What happened: NVIDIA introduced Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise databases with no task-specific training, feature engineering, or table flattening.
Key numbers: 41 tasks across 15 databases; up to 10,000 context examples per query; 8% MRR gain over AutoGluon on SAP SALT; nearly 4 AUROC points over
Why it matters: NVIDIA states it is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks, and argues the results make a case for replacing production XGBoost pipelines.
What changed
NVIDIA introduced Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise databases with no task-specific training, feature engineering, or table flattening.
Why it matters
NVIDIA states it is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks, and argues the results make a case for replacing production XGBoost pipelines.