Attentive State Space Modeling of Dynamical Systems
Abstract
In this project, we develop a new class of deep probabilistic models for dynamical systems, such as surveillance systems and disease progression. The proposed model will be capable of informing decision making by learning highly accurate representations of dynamical systems, and extracting intelligible knowledge from data. The model will achieve this by combining the complementary strengths of deep learning, which is capable of learning flexible and accurate data representations, and probabilistic state-space models, which are capable of distilling intelligible structure that explain a system~s dynamics. Our deep probabilistic framework for dynamical systems capitalizes on both the principled interpretable representations of probabilistic models and the predictive strength of deep learning methods. The proposed model will use a state-space representation to segment a dynamic trajectory into ~stages~ of progression that manifest through observations. But unlike conventional state-space models, which are predominantly Markovian, our model will use recurrent neural networks (RNN) to capture more complex state dynamics.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N000141912510
Entities
People
- Mihaela Van Der Schaar
Organizations
- Office of Naval Research
- United States Navy
- University of Cambridge