Relational data without training
NVIDIA introduced Kumo Relational, a model that works directly on the links between tables in relational databases. The company claims it beats conventional supervised machine learning approaches without any training step.
Relational data refers to tables of rows and columns connected to each other through key fields. Much of banking, e-commerce and enterprise record-keeping is stored this way.
Why it matters
Classically, producing predictions from this kind of data requires feature engineering and model training. Removing the training step could let organizations get results faster from data they already hold.
The announcement follows NVIDIA's recent expansion of its open model family for Europe, which also introduced Kumo. The company is thus extending its AI models beyond hardware into the enterprise data layer.
