-
Notifications
You must be signed in to change notification settings - Fork 0
Home
BIDMat is a matrix/machine learning toolkit designed to be the fastest and most scalable solution for large-scale behavioral data analysis on single machines or clusters. In more detail its designed to:
1. Provide an interactive data analysis environment, similar to R or Matlab, for exploratory modeling and analysis, but with a fast deployment path to distributed implementation.
2. Leverage CPU performance through native libraries (Intel MKL, HDF5, string/XML processing).
3. Leverage GPU hardware and GPU-based data as first-class citizens.
4. Provide high-speed, low-latency communication primitives on clusters (butterfly mixing) for the fast model-mixing needed by stochastic gradient and monte-carlo optimization methods.
5. Leverage recently-developed block linear-systems and eigenvalue solvers. These solvers use blocks of vectors or multi-vectors to reduce the number of passes over large datasets when solving linear equations or computing eigenvalues.