Nvidia details Kumo Relational model and claims gains on database prediction tests

Nvidia introduced Kumo Relational on October 21, describing a pretrained model that predicts outcomes from linked database tables. The company says it can tackle questions about customer departures, demand and fraud without training a separate model for each task.

Nvidia evaluated the model across 41 tasks involving 15 databases. On SAP SALT, it reported an 8% advantage over AutoGluon ensembles on the mean reciprocal rank measure. These are results presented by Nvidia, which said the evaluation datasets were excluded from pretraining.

The model draws on as many as 10,000 historical examples when generating predictions, preserving connections between tables. That approach reduces the need to manually combine records and design predictive inputs. Nvidia acknowledged that specialized models can still perform better when very large training datasets are available.

Developers can access Kumo Relational through Nvidia's structured-data-models library or deploy it through Nvidia NIM. Further tuning on a particular dataset is also available for applications needing better accuracy.