Use Cases

Experiment/model sharing

OpenML.org store many ML experiments, including models and full predictions. It would be great to export these as linked open data, and connect them to other uses/knowledge of the datasets and algorithms. This will also facilitate the sharing of experiments across tools. OpenML also has lots of metadata on the algorithms and datasets that would be useful to export.

Caffe has a collection models, called modelzoo. Describing these will be very useful for discovering useful deep learning models.

Workflow design/planning

Designing workflows for machine learning is a difficult problem. Proper annotations of workflow operators and the outcomes of entire workflows are useful to further research in this direction.

Meta learning

Use large amounts of shared experiments to understand when algorithms work very well or not, or how to select and use them given a certain dataset. This is related to the previous point.

Text mining

Do text mining on machine learning articles, extract knowledge, and express it uniformly (information extraction - IE)

Experiment reproducibility in publications

Publications could be extended by a proper description of the involved experiments. This will allow clearer interpretation and better reproducibility.

Research Objects and ROHub portal are working in this direction and could be interested in ML-Schema.

Comparison of ML algorithms.

There are different criteria for the selection of algorithms and for the comparison of their performances. But it is not clear in what situation what criteria to select. Such an analysis would require a crisp description of algorithms and their results on particular datasets.

For example, when working with prediction of drug activities, the normal criteria often are not suitable. Researchers do not care about the overall performance, but only about the performance regarding the few top predictions.

##Education Using ML Schema for defining learning objects which contain definitions of typical terms in the domain.