Supercomputing and Computation
Ultrascale Predictive Analysis ToolboxComputational Data Analytics Group
May 14, 2013
- Supercomputing’s major successes to date have come from discovery through simulation rather than discovery through analysis
- Traditional multivariate analysis is not scalable to extremes; high-dimensional data are broken into subsets that ignore system-level context
- Can we tractably analyze ultrascale complex-systems data, learning from the system-wide interactions?
- Use system-level interaction information to address the sample complexity problem.
- Treat variables as tasks in a multi-task machine learning approach.
- Provide the capability for scientists to explore ultrascale datasets in new ways.
Advantage over the State-of-the-Art:
- Rather than being stymied by complexity, our approach takes advantage of complex interdependencies in a system to learn in ways that aren’t possible in less complex systems
- Allows data from different time periods, different locations, and even different variables to be used together without typical fusion problems, such as data imputation for missing variables.