IEEE - Aerospace and Electronic Systems - July 2021 - 91

Blasch et al.
Figure 13.
AI/ML and SDF integration ofedge and cloud information.
changes include command by intent, command by plan,
and command by control. Each command approach
requires data, sensor, information, and intelligence fusion
for heterogeneous coordinated asset control to empower
local initiative within the commander's mission intent [33].
EXAMPLE: EDGE ANALYTICS
AI/ML integration of data from the surveillance sources as
well as edge devices helps to determine the situation. The targeting/decision
options are: detect objects, disable communication,
discern behavior, dissuade harmful results, and
determine situational actions. In a complex battlefield, integrating
data-centric with model-centric ML SDF combines
strategic surveillance with tactical human sensor knowledge.
Leveraging ofthe models ofthe underlying physical phenomena
supports explainable solutions for situation knowledge.
Figure 13 demonstrates how an edge device with cameras and
wireless connections in a DL approach can serve to provide
data to another DL layer to integrate cloud data towards a
labeled decision. Figure 13 leverages the traditional CNN by
first processing the distributed edge devices and then aligns
the results for the traditional three layers ofthe CNN: convolution
(to learn features), max polling (as dimension reduction),
and fully connected layer (for classification).
The basis of information fusion is uncertainty analysis.
An AI/ML approach with SDF offers reliable error bounds
to discriminate the target type. A concern for AI/ML is how
to continuously evolve/adapt intelligent fusion components
in an ever-changing environment without the benefit oflarge
amounts of labeled data. An example is a GAN that generates
model-based data that when combined with collected
data allows both machine and human discrimination through
a SenseGAN [34] for robust situation analysis. For regression
problems, the area between the calibration curve of an
algorithm and the optimal calibration curve, called deviation
area, represents the measurement of uncertainty as the
smaller the deviation area is, the better the algorithm estimates
the quality of uncertainty. Figure 14 further explores
the energy needed, as determined from the number of learning
iterations, for IoBT sensing devices towards decision
making. The RDeepSense [35,36] provides an efficient and
effective method demonstrating the benefits of AI/ML with
SDF while increasing object assessment. Compared with the
simple and scalable predictive uncertainty estimation algorithm,
RDeepSense uses dropout regularization as an
implicit ensemble method, which avoids running multiple
DL models with Monte Carlo sampling methods such as
dropout duringmodel inference on embedded devices.
CHALLENGES AND OPPORTUNITIES
The challenges include the optimal distribution of
resources for SDF, training intelligent sensors with little
or no labeled data, and uncertainty analysis. For example,
there is a concern on how to optimally place AI components
in the distributed (likely contested) edge system
for maximum fusion speed and efficacy and for minimum
communication/emission footprint. The system must
automatically distribute prediction and inference computations
over heterogeneous device networks with varying
connectivity and bandwidth. Such an example is to map
a multilayer prediction and inference architecture onto a
physical device network as well as jointly optimize all
layers to account for device resource limitations,
expected network conditions, and stealthy requirements.
The coordination of AI/ML with SDF provides many
Figure 14.
RDeepSense providing optimal results with minimal energy.
JULY 2021
examples and research opportunities to exploit GANs to
generate labeled data samples to train for operational
needs. For example, dropout training in DNN and
IEEE A&E SYSTEMS MAGAZINE
91
IEEE - Aerospace and Electronic Systems - July 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2021
