Galois is innovating ways to create accessible data models or make existing data models actionable during a crisis like COVID 19. I blogged about this in April and, more specifically, in October when I wrote about our AMIDOL project.
This month, I wanted to talk about our exciting work with SuperMaaS – a modeling-as-a-service (MaaS) model registry framework for existing models.
Galois is developing the SuperMaaS project as part of DARPA’s World Modelers program. Galois is also partnering with the Bill and Melinda Gates Foundation and the Ethiopian government’s Decision Support Modeling Tools for Ethiopia (DSMT-E) project, which focuses on food security modeling.
In September 2020, DARPA and stakeholders from DSMT-E were able to use a simulated version of SuperMaaS. The simulation involved “food shock cascades.” A “food shock” is any crisis that affects scarce food supply. A food shock could be a natural disaster, or it could result from a societal effect, such as a spike in agricultural prices.
Participants were able to model the effects of the following food shock cascades on Ethiopia’s food supply: COVID 19’s supply chain disruptions, desert locusts affecting crops; floods; and refugees migrating from other countries.
I’m pleased to say that SuperMaaS worked as we intended: participants successfully examined data from each food shock cascade and predicted outcomes using the SuperMaaS tool.
But how, exactly, does SuperMaaS work? Read on to learn more.
“Look! Up there in the cloud! It’s … SuperMaaS!”
Modeling-as-a-Service (MaaS) is a state-of-the-art solution for outputs based on existing data sets. SuperMaaS takes this concept further. Our idea is to combine human and machine intelligence so that professionals of all skill levels – domain modelers, general modelers, and policymakers – can update a data model in real-time, with fewer errors. Many of us “team-up” with machine learning software on a daily basis; if you’re typing on a word processor, you’re teaming up with a machine. SuperMaas is designed to do a far more complicated task: multiple groups teaming up with multiple machines. Experts would use interfaces designed specifically for their work. For example, an analyst’s interface would look different from a general modeler’s interface (which can see the underlying model) and a domain modeler’s interface.
The Justice League of Data Models
We intend to make SuperMaaS a model registry for domain modelers, but how do we do that? This is the process we are developing:
- The goal of the model registry process is to simplify many tasks that domain modelers do by hand. 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. This would work for hand-coded models, machine-learned models, and even models that are nothing more than simple scripts.
- SuperMaaS uses containers, instead of virtual machines, to create interfaces. Containers provide powerful capabilities to handle complex run-time requirements, execution environments, and more.
- Our end goal is 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.
Fortress of Datacubes
SuperMaaS also seeks to simplify model outputs. Currently, model outputs are not actionable in real-time because they are often incompatible with other models. SuperMaaS is testing a new concept called the datacube, which is a type of point lattice. The axes of the datacube would be annotated with mappings to a hierarchical concept graph within SuperMaaS. This would help users search for, interpret, and transform datacubes for analysis, visualization, and comparison.
Datacubes produced by models will also be labeled automatically, allowing users to easily search for different types of data. For instance, 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 to see.
Additionally, the general modeler interface in SuperMaaS helps a general modeler discover models or data within SuperMaaS that can be used to generate datacubes based on requests from analysts.
Real-World Problems Solved in Real-Time
So, what’s next for SuperMaaS? Participants at our September presentation helped identify some bugs that we are currently addressing. For example, we will work to smooth out interactions between analysts, general modelers, and domain modelers and close feedback loops.
Data scientists, academics, and policymakers are the real-life superheroes of our world. And the effort to amass real-world models in real-time and then update those same models is nothing short of superheroic.
Stay tuned for further updates, true believers!
To read more about SuperMaaS, see our project page.
You can also read about the SuperMaaS Hackathon event we held in June 2020.
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.