+1
See also this thread.
Main points: via deeplearning4j knime already has nd4j available which is basically a numpy for java. A first step for me would be a new java snippet node that can operate on the whole table + also has nd4j or similar lib available for fast operations.
As I mentioned in the linked thread, for me it’s not strictly about classic math matrix operations but being able to work on the whole table within java /JVM so not having to pay the serialization penalty which sadly is huge. The combined time of serialization/loading results usually takes more time than the actual calculations.
EDIT: Note that nd4j has gpu (CUDA) support built-in. I mean being able to perform gpu accelerated matrix operations in a GUI / for non-programmers could be a huge selling point. Just saying…