Developing New Simulation Models for Machine Learning
Abstract
Although machine learning techniques have enjoyed recent successes for varied engineering problems, they also suffer from a number of issues which researchers have worked to address in often disparate ways. We feel that introducing an underlying structural and first-principles based model has the potential to ground the machine learning paradigm so that it introduces data as an augmentation to underlying popular and existing approaches as opposed to being saddled with the task of understanding and interpreting the data with little context. Examples of marrying models and data have existed for some time, such as the symbiotic relationship between computer vision and computational geometry. In fact, the quite popular method of snakes (active contours) mixes physical simulation models with the dataanalysis and interpretation. This, i.e. leveraging physical simulation, is what we propose to do here as well; however, we will focus on problems related to humans such as machine learning approaches for human-robot teaming and task/activity identification. We will initially take a model heavy physical simulation approach towards embracing the data, since our lab has extensive expertise in unique innovative physical simulation and we believe not only that a novel numerical approach will be required but also that few research groups will be able to explore such an avenue - i.e., we are not proposing that this is obviously the best and most profitable direction, but we are confident that it willplay a role and thus expertise such as ours will be required.We will build high-fidelity research techniques that would for example allow the simulation of detailed facial expressions as well as the form and motion of the human body and clothing (including uniforms and gear). For some applications, where this is overkill, we anticipate that similar ideas could be used on lower-fidelity, more computationally efficient models. This same trickledown approach has worked well in moving various technologies from feature film production toreal-time environments including augmented/virtual reality and games. The utilization of data with the statistical and machine learning techniques will augment this transition back and forth between more and less detailed physical models as the data allows us to replace and augment certain aspects of the physical model depending on the situation and requirements such as certain forms of accuracy. However, history has shown that first exploring things in great detail in order to obtain a full understanding of the task at hand quite often leads to largely unanticipated futuresimplifications which often enjoy efficiency and accuracy at a level unattainable by some of the more straightforward adhoc approaches. Thus, the main goal of our project is to explore and begin to understand what the approach to physical simulation should even be when the eventual goal is to tightly intertwine physical simulations and data. We are uniquely qualified to do this since we did it once before leading the charge to modify the entire process of physical simulation into something different when it came to Hollywood special effects and computer graphics as compared tophysical simulation for computational physics.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712174
Entities
People
- Ronald Fedkiw
Organizations
- Office of Naval Research
- Stanford University
- United States Navy