Publications

Recent works in deep learning have been driven broadly by the desire to attain high accuracy on certain challenge problems. The network architecture and other hyperparameters of many published models are typically chosen by trial-and-error experiments with little considerations paid to resource constraints at deployment time. We propose a fully automated model learning approach that (1) treats architecture selection as part of the learning process, (2) uses a blend of broad-based random sampling and adaptive iterative refinement to explore the solution space, (3) performs optimization subject to given memory and computational constraints imposed by target deployment scenarios, and (4) is scalable and can use only a practically small number of GPUs for training. We present results that show graceful model degradation under strict resource constraints for object classification problems using CIFAR-10 in our experiments. We also discuss future work in further extending the approach.

Ground personnel at the tactical edge often lack data and analytics that would increase their effectiveness. To address this problem, this work investigates methods to deploy cloud computing capabilities in tactical environments. Our approach is to identify representative applications and to design a system that spans the software/hardware stack to support such applications while optimizing the use of scarce resources. This paper presents our high-level design and the results of initial experiments that indicate the validity of our approach.

As new observation systems are developed and deployed, new and presumably more precise information is becoming available for weather forecasting and climate monitoring. To take advantage of these new observations, it is desirable to have schemes to accurately retrieve the information before statistical analyses are performed so that statistical computation can be more effectively used where it is needed most. The authors propose a sequential variational approach that possesses advantages of both a standard statistical analysis [such as with a three-dimensional variational data assimilation (3DVAR) or Kalman filter] and a traditional objective analysis (such as the Barnes analysis). The sequential variational analysis is multiscale, inhomogeneous, anisotropic, and temporally consistent, as shown by an idealized test case and observational datasets in this study. The real data cases include applications in two-dimensional and three-dimensional space and time for storm outflow boundary detection (surface application) and hurricane data assimilation (three-dimensional space application). Implemented using a multigrid technique, this sequential variational approach is a very efficient data assimilation method.

Air traffic congestion caused by convective weather in the US has become a serious national problem. Several studies have shown that there is a critical need for timely, reliable and high quality forecasts of precipitation and echo tops with forecast time horizons of up to 12 hours in order to predict airspace capacity (Robinson et al. 2008, Evans et al. 2006 and FAA REDAC Report 2007). Yet, there are currently several forecast systems available to strategic planners across the National Airspace System (NAS) that are not fully meeting Air Traffic Management (ATM) needs. Furthermore, the use of many forecasting systems increases the potential for conflicting information in the planning process, which can cause situational awareness problems between operational facilities. One of the goals of the Next Generation Air Transportation System (NextGen) is to consolidate these redundant and sometimes conflicting forecast systems into a Single Authoritative Source (SAS) for aviation uses. The FAA initiated an effort to begin consolidating these systems in 2006, which led to the establishment of a collaboration between MIT Lincoln Laboratory (MIT LL), the National Center for Atmospheric Research (NCAR) Research Applications Laboratory (RAL), the NOAA Earth Systems Research Laboratory (ESRL) Global Systems Division (GSD) and NASA, called the Consolidated Storm Prediction for Aviation (CoSPA; Wolfson et al. 2008). The on-going collaboration is structured to leverage the expertise and technologies of each laboratory to build a CoSPA forecast capability that not only exceeds all current operational forecast capabilities and skill, but that provides enough resolution and skill to meet the demands of the envisioned NextGen decision support technology. The current CoSPA prototype for 0-6 hour forecasts is planned for operation as part of the NextGen Initial Operational Capability (IOC) in 2013. CoSPA is funded under the FAA's Aviation Weather Research Program (AWRP). The first CoSPA research prototype demonstration was conducted during the summer of 2008. Technologies from the Corridor Integrated Weather System (CIWS; Evans and Ducot 2006), National Convective Weather Forecast (NCWF; Megenhardt et al. 2004), and NOAA’s Rapid Update Cycle (RUC; Benjamin et al. 2004) and High Resolution Rapid Refresh (HRRR; Benjamin et al. 2009) models were consolidated along with new technologies into a single high-resolution forecast and display system. Historically, forecasts based on heuristics and extrapolation have performed well in the 0-2 hour window, whereas forecasts based on Numerical Weather Prediction (NWP) models have shown better performance than heuristics past 3-4 hours (Figure 1). One of the goals of CoSPA is to optimally blend heuristics and NWP models into a unified set of aviation-specific storm forecast products with the best overall performance possible. The CoSPA prototype demonstration began in July 2008 with 2-6 hr forecasts of Vertically-Integrated Liquid water (VIL) that seamlessly matched with the 0-2 hr VIL forecasts available in CIWS. These real-time forecasts have been made available to the research team and FAA management only through a web-based interface. This paper discusses the system infrastructure, the forecast display, the forecast technology and performance of the 2-6 hr VIL forecast. Our early assessment based on the 2008 demonstration is that CoSPA is showing tremendous promise for greatly improving strategic storm forecasts for the NAS. Early user feedback during CoSPA briefings suggested that the 6 hr forecast time horizon be extended to 8 hours to better meet their planning functions, and that forecasts of Echo Tops must also be included.