Human Guided Unified Learning of Tractable Deep Probabilistic Models

Abstract

This project seeks to develop unified, efficient, effective and human-guided algorithms for learning Tractable Deep Probabilistic Models (TDPMs). TDPMs combine the efficiency, ability to handle latent information and ease of handling hybrid data of deep models with the tractability, interpretability and ability to handle structured domains of probabilistic graphical models. We propose to first develop formalisms for unifying different data types, domains types and model types. Next, we propose effective learning methods for both discriminative and generative TDPMs based on gradient-boosting. Then, we extend these algorithms to handle rich human guidance in the form of qualitative constraints and preferences. Finally, we propose to develop automated statistician in medical domains where the algorithms will be rigorously evaluated. The transformative nature of the proposal allows effective and efficient learning of Tractable Deep Probabilistic Models effectively with human guidance. If successful this proposal has the potential to advance research in multiple areas - TDPMs can be natural function approximators for Reinforcement Learning (RL) and learning TDPMs with human guidance can realize the grand vision of ``safe RL". Our proposed unifying formalism of Kolmogorov-Lorentz Networks can be very effective due to the resultant sum-sum substructures. Finally, given that one could read conditional independencies directly off of an SPN, our TDPMs can realize the grand vision of an automated statistician where a data set can be given as an input and with some human guidance, the learning system can automatically read and create a notebook of conditional and context-specific independencies that exist in the data. A key application of the proposed framework is in understanding the authorship of ICML papers. The resulting algorithms from this framework will enable the development of comprehensive models of research areas. For instance, one could infer how the field has evolved, what are the most promising ideas, how the inspirations have shaped the field etc. Combined with our evaluations on a road network domain, these ideas can be extended to understanding the structure of the urban environments. For instance, one could understand the movements of the people, where they work and where they live in. One could understand which neighborhood needs additional resources (in case of insurgency). The ideas developed in this proposal can allow for an unified learning framework that can learn hybrid probabilistic deep models at different levels of abstraction.

Document Details

Document Type
DoD Grant Award
Publication Date
Jul 09, 2020
Source ID
W911NF2010224

Entities

People

  • Sriraam Natarajan

Organizations

  • Army Contracting Command
  • United States Army
  • University of Texas at Dallas

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks