PROBABILISTIC PROGRAMMING FOR ADVANCED MACHINE LEARNING (PPAML) DISCRIMINATIVE LEARNING FOR GENERATIVE TASKS (DILIGENT)
Abstract
This Final Report summarizes the research conducted in the course of DARPA PPAML program, which was focused on enabling the use of discriminative solvers to solve discriminative tasks specified in probabilistic programs. The research produced two complementary methods of constructive discriminative solvers: model-driven and data-driven ones: research frameworks and software implementations) in both generative and discriminative areas: Generative models: a novel framework for most accurate computation of key statistical elements of model-driven problems (such as conditional probability, regression, etc.) Discriminative models: a novel framework for capturing domain knowledge in the form of features and kernels for standard data-driven problems (solved in LUPI approaches). These achievements are described in in this report and in 9 papers published in the course of DARPA PPAML program.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 29, 2017
- Accession Number
- AD1042315
Entities
People
- Rauf Izmailov
- Vladimir Vapnik