IEEE - Aerospace and Electronic Systems - July 2021 - 72

Artificial Intelligence and Data Fusion at the Edge
Table 3.
ML Aligned With IF
ML Methods
Registration,
Estimation
CNN, RNN, LSTM,
Estimation
CNN, Pattern
Matching
GAN
RL, Optimization,
Regularization
RL
*
(Bayes)
@
(DL)
*
Types of ML
Symbolic Probabilistic Connectionist Analogistic Evolutionary Possibilistic
(Logic)
(SVM)
(GA)
(Fuzzy)
@ ** L0 Data
Assessment
* @@
@
@
@
@@ @ ** @
* @@ *
@
*
* @@ ** @
Active Learning @@ @ ** *
@@ @ ** *
L1 Object
Assessment
L2 Situation
Assessment
L3 Impact
Assessment
L4 Process
Refinement
L5 User
Refinement
L6 Mission
Refinement
connected handwriting recognition and speech recognition.
LSTM is a type of RNNs that have feedback
connections and help overcome the vanishing
gradient problem in RNNs. GAN is an ML framework
based on ANNs that learns to generate new
data with the same statistics as the training set. RL
is another area of ML that is concerned with determining
an optimal set of actions in an environment
in order to maximize cumulative reward. RL is typically
stated in the form of a Markov decision process
(MDP), where dynamic programming provides
a solution for the MDP [28]. DRL combines DNNs
with RL algorithms (e.g., Q-learning) to solve previously
unsolvable problems as DRL can learn from
raw sensor data and/or images supplied as input. In
active learning, an ML algorithm can interactively
pose queries during the training process, usually in
the form of unlabeled data instances to be labeled
by a human user. Thus, active learning is an example
of human-in-the-loop learning. Connectionist
ML often leverages regularization. Regularization
is a technique that makes slight modifications to the
ML algorithm to prevent overfitting to the training
data and to make the model more generalizable for
different test datasets. Connectionist ML also preprocesses
the input data through data registration.
Registration, often performed for images and thus
known as image registration, is the process of transforming
different images of one scene into the same
coordinate system. Registration is required in order
to compare or fuse the data obtained from different
measurements. For example, images can be taken at
different times (multitemporal registration), by
72
different sensors (multimodal registration), and/or
from different viewpoints [29].
(4) Analogistic ML. Analogistic ML analyzes data analogies
and similarities through distance computations
in feature hyperspace. Examples of analogistic
ML include support vector machines (SVMs) and
nearest neighbors, such as K-nearest neighbor
algorithm.
(5) Evolutionary ML. Evolutionary ML utilizes computational
models inspired by evolutionary competition
and survival. Examples of evolutionary ML
include genetic algorithms (GAs), genetic programming,
and neuroevolution. We note that neuroevolution
is similar to genetic programming but the
genomes represent ANNs by specifying structure
and connection weights.
(6) Possibilistic ML. Possibilistic ML analyzes ambiguous
data using extension of classical logic to represent
partial truths. Fuzzy inference systems and
possibilistic logic systems are examples of possibilistic
ML.
ADVANTAGES OF DATA FUSION AND AI AT THE EDGE
Data fusion and AI at the edge can provide various advantages
in terms of latency, energy, accuracy, security, privacy,
cost, scalability, and sustainability as discussed in
the following.
Latency. Latency refers to the time spent in the whole
AI inference process, including pre-processing, data
fusion, model inference, data transmission and post
IEEE A&E SYSTEMS MAGAZINE
JULY 2021
Information
Fusion
Alignment
IEEE - Aerospace and Electronic Systems - July 2021

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