SuperMaaS: A Proposed Model Registry For Actionable Updates

This project, part of DARPA’s World Modelers program, seeks to create a model registry that combines human and machine intelligence for iterative and actionable updates in real-time.

Policymaking during a crisis is experiencing its own crisis. Academics, policymakers, and researchers all work separately on a data model, limiting any ability to make real-time updates. Modeling-as-a-Service (MaaS) has emerged as a method to address these foundational problems, and Galois is working on a project that could advance this technology. 

Galois is developing the SuperMaaS project as part of DARPA’s World Modelers program. Galois’s goal is to help domain experts create models from data sets. 

A Real-Time Model Registry 

The SuperMaaS framework plans to create a model registry for domain modelers. The model registry would take a domain modeler’s code and create an abstract version of the model in SuperMaaS. The model would then be considered a registered model in SuperMaaS. 

The model registry takes common design patterns present in many scientific models and creates a simplified interface for all registered models residing in SuperMaaS. This would work even for different types of models, including hand-coded models, machine-learned models, and even models that are nothing more than simple scripts. The goal of the model registry process is to simplify many tasks that domain modelers do by hand. 

Galois aims to do this by using containers, a “lighter-weight” option than virtual machines. Containers still provide powerful capabilities to handle complex run-time requirements, execution environments, and more.

The SuperMaaS registry process will employ a domain-specific registry language which will let domain experts define their modeling environments, execution procedures for outputs, how to capture and show outputs, and a way to show modelers how to render their inputs. 

Datacubes Simplify Model Outputs

SuperMaaS is also intended to simplify model outputs. Currently, model outputs are highly complex and are not actionable in real-time because they are not compatible with other models. SuperMaaS is testing a new concept called datacubes to create standardized schema. Galois has been conceptualizing datacubes as a type of point lattice. A point lattice defining a given datacube has axes corresponding to the exposed inputs of the registered model that generated it.

The axes of the point lattice would be annotated with mappings to a hierarchical concept graph within SuperMaaS, which will help users search for, interpret, and transform datacubes for analysis, visualization, and comparison.

Galois is currently working to address multiple names for each concept so that both the concepts and the names can be easily connected. The University of Arizona is assisting Galois in collecting and organizing concept datasets into a usable hierarchy of related terms using their Eidos, INDRA, and Delphi technologies. 

Once in the SuperMaaS system, these concepts will be stored in hierarchical maps that

connect to models, indicating the role they play and the concepts used by the model.

This will help users find relevant inputs and datasets, and it will also help them interpret results.

Datacubes produced by models will also be labeled automatically, allowing users to easily search for different types of data. For example, if a user had a datacube with a temperature in Kelvin but needed data representing Celsius, SuperMaaS would locate the correct transforms so they can find the part of the concept map they need.

Actionable Updates in Real-Time

The eventual goal of the SuperMaaS framework is to make the process of building a registered model from its components as easy as possible. Galois aims to support a registry process that can be completed in a day or less by a domain expert, with minimal knowledge of the process of containerization. 

You can read about our latest work with SuperMaaS on our blog.

You can also read about our June 2020 SuperMaaS hackathon here.

This material is based upon work supported by the Army Research Office/Defense Advanced Research Projects Agency and the Army Contract Command under Contract No. W911NF-20-C-0056.

Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Army Research Office/Defense Advanced Research Projects Agency and the Army Contracting Command.