Hybrid Learning: Algorithm and Architecture for Coupling Model and Data to Learn Dynamics of Complex Systems (Topic Area:Algorithms and Architectures)
Abstract
Computing Grand Challenge: Learning and predicting behavior of complex dynamical systems, where perfect or complete information of the state and/or dynamics is not known, pose a major challenge to cognitive computing. Many real-life applications, such as weather forecasting, battlefield decision-making, autonomous vehicles, swarm robotics, traffic monitoring, are examples of such complex dynamic environments. Fueled by the advancements in data-driven deep learning, today s artificial intelligent (Al) systems have made significant progress in pattern recognition and classifications; but deep learning has limited ability to understand the internal dynamics of a complex system. On the other hand, physical models of dynamical systems often suffer from lack of exact knowledge, the uncertainties in model parameters and measurements, and inability to capture hidden dynamics in the system. Therefore, a grand challenge for future intelligent computing platforms is to be able to learn dynamics of a complex system Objectives: The proposed research seeks to develop an AI model, referred to as hybrid learning, that interacts with a complex system to learn its dynamics, generates hypothesis of its behavior, and makes prediction. The hybrid learning model will integrate statistical learning, dynamical system models, and neuro-inspired learning to enable future computing systems perform causal reasoning, inference and discovery. The proposed research will explore algorithm-architecture co-design to achieve more than lOOX gains power-efficiency of computing platforms for hybrid learning. Intellectual Merit The proposed research couples data-driven statistical machine learning techniques, dynamical system based model-driven computation, and neuro-inspired learning in a single architecture to predict evolution of coupled system with hidden dynamics. The developed hybrid network learns to predict system evolution solely from observable states, including direct, noisy, and encoded observations. This will require modeling known dynamics, discover hidden dynamics from observations, and couple them in a cohesive manner in a single network. The hybrid network accurately predicts state evolution even if the relational parameters, physical properties, or composition of complex system change. This is achieved by continuously learning physical parameters and/or governing dynamics from observations. Our model discovers causal relations, rather than correlated relations, from observations to support causal reasoning, inference, and prediction of states and events. The spiking neural network (SNN) with stochastic spike-time-dependent-plasticity (STOP) based learning is integrated within the hybrid network to enable causal reasoning and prediction of events. Finally, the proposed model demonstrates high power-efficiency while solving real-world problem in diverse application domains and efficiently run on computing platforms with varying resources. Significance: The proposed research will fundamentally advance capabilities and power-efficiency of future computing platform interacting with complex systems. Learning and discovering dynamics of complex systems via integration of data- and model?driven computation will fundamentally advance analysis, prediction, and decision-making capabilities of computing platforms operating in dynamic environments. Co-design of learning algorithm and hardware architecture will instigate lOOX gains in power-efficiency to enable application of hybrid learning in both high-performance and resource-constrained environments in many DoD and commercial applications, including autonomous systems, unmanned aerial vehicles, swarm robotics, battlefield decision making, network analysis, and scientific computation, to name a few.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 01, 2019
- Source ID
- W911NF1910477
Entities
People
- Saibal Mukhopadhyay
Organizations
- Army Contracting Command
- Georgia Tech Research Corporation
- National Security Agency