NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

NVIDIA Kumo Relational demonstrates superior performance over supervised machine learning when handling relational data, notably without requiring explicit training. The model operates by taking in-context training tables (x_context, y_context) and prediction examples (x_query) as sdm.TableTensor objects. It also utilizes multi-table information for both context and prediction entities via sdm.RelatedTables. Users can deploy Kumo Relational as an NVIDIA NIM service and interact with it using the open-source structured-data-models-client, writing predictions in PQL.

This report highlights NVIDIA Kumo Relational's unique capability to outperform supervised machine learning on relational data without any training, unlike traditional models that require extensive pre-training.