PROSPERO: Provenance Supporting Pictorial Model Building

SIFT and Raytheon/BBN propose PROSPERO: Provenance Supporting Pictorial Model Building to create tools to assist in dataset and model curation for machine learning. PROSPERO will support an interactive Workspace that will allow model builders to visually inspect the provenance of any dataset or model. áPROSPERO will import data into a knowledge graph that will support automated reasoning to enforce quality constraints in model building. á PROSPERO will also support reasoning about the types of data allowing an organization to make custom datasets suitable for sharing with other organizations that do not reveal their sensitive information. PROSPERO workspace has two modes; building and quality exploration. PROSPERO builds new datasets by allowing the user to explore existing datasets and combine or transform the data to produce a new dataset of known quality for a target use. For quality exploration PROSPERO provides a visual provenance graph that chronicles the sources of the data and every transform applied to the data. The interactive provenance graph precisely describes the steps taken to build a dataset or model. PROSPERO will support; provenance graphs the show the source and all transformation applied to a data set or model, integrity checking plug-ins that ensure the knowledge graph has specific properties, tear-lines controlled by an organization to áprevent their sensitive information from being shared, a visual Workspace for inspecting and building datasets. áPROSPERO also supports an API to allow other components of the TAK ecosystem add provenance information for the actions they take on datasets or models. PROSPERO system to allow novice and expert users to efficiently produce new high quality datasets and customize existing ML models to enable TAK to provide exciting new ML driven capabilities. á