2 Problem Analysis
Generally speaking, the main problems of the application of AI technology in military
are as follows:
First is data. Data is the lifeblood of AI. Without dataset, AI has become a passive
water. Learning has neither object nor verification basis. Data problems include three
aspects: Firstly, the problem of data source—in peacetime, operational data mainly
comes from the aggregate of daily duty, live exercise, training and other ways.
However, these aggregates are far from enough to effectively support machine
learning. Even the USA, which has fought the most war in the world in recent years,
has the richest aggregate of battlefield data in the world, but its Defense Innovation
Unit (DIU) still thinks that the key to restricting the application of AI in the military is
that there are too few machine learning datasets available in a short and strong
competed environment, which makes it difficult for AI systems to carry out tasks
effectively in the complicated and competed environment [1]; secondly, the problem
of data comprehensiveness—the rapid development of AI technology is mainly due to
the continuous development of deep learning system. By deep learning, a large
amount of information can be inputted into the computer to allow deep learning
system to learn and analyze data. However, the data used to train deep learning system
is not comprehensive although it is constantly improving. For example, when using AI
system to pick out fruits, people train visual recognition algorithms by learning a large
number of image databases. The purpose is usually to identify the “natural” appearance of fruits. However, compared with the bright photographs of intact fruits, there
are relatively few pictures of rotten fruits. And unlike humans, AI systems tend not to
calculate or ignore them, while the human brain tends to pay attention to these
abnormal groups and react strongly to them [2]. In military application scenarios, this
situation is even more serious; and thirdly, the problem of data annotation—current
machine learning systems acquire a large amount of data through sample learning,
which has been independently labeled by analysts to generate the required output.
With the development of the system, deep neural network (DNN) has become the
latest technology in machine learning model. DNN can provide power for tasks such
as machine translation and speech or object recognition with higher accuracy. However, training DNN requires a lot of labeling data. The process of accumulating and
labeling a large amount of information is expensive and time-consuming. In addition
to the challenge of accumulating labeled data, most machine learning models are
fragile and vulnerable to collapse when their operating environments change slightly.
For example, in speech recognition systems, if the acoustic environment or the sensor
of the microphone changes, it may be necessary to retrain on a completely new
dataset. The time and effort required to adjust or modify the model are almost
equivalent to recreating the model.
Second is the problem of autonomous learning. At present, AI takes machine
learning technology as its core and relies on huge dataset. However, the battlefield
environment is complicated and changeable, the competition is violent, the situation is
changing rapidly, full of fog and accident, and the AI system cannot predict all possible
408
G. Zhou
Generally speaking, the main problems of the application of AI technology in military
are as follows:
First is data. Data is the lifeblood of AI. Without dataset, AI has become a passive
water. Learning has neither object nor verification basis. Data problems include three
aspects: Firstly, the problem of data source—in peacetime, operational data mainly
comes from the aggregate of daily duty, live exercise, training and other ways.
However, these aggregates are far from enough to effectively support machine
learning. Even the USA, which has fought the most war in the world in recent years,
has the richest aggregate of battlefield data in the world, but its Defense Innovation
Unit (DIU) still thinks that the key to restricting the application of AI in the military is
that there are too few machine learning datasets available in a short and strong
competed environment, which makes it difficult for AI systems to carry out tasks
effectively in the complicated and competed environment [1]; secondly, the problem
of data comprehensiveness—the rapid development of AI technology is mainly due to
the continuous development of deep learning system. By deep learning, a large
amount of information can be inputted into the computer to allow deep learning
system to learn and analyze data. However, the data used to train deep learning system
is not comprehensive although it is constantly improving. For example, when using AI
system to pick out fruits, people train visual recognition algorithms by learning a large
number of image databases. The purpose is usually to identify the “natural” appearance of fruits. However, compared with the bright photographs of intact fruits, there
are relatively few pictures of rotten fruits. And unlike humans, AI systems tend not to
calculate or ignore them, while the human brain tends to pay attention to these
abnormal groups and react strongly to them [2]. In military application scenarios, this
situation is even more serious; and thirdly, the problem of data annotation—current
machine learning systems acquire a large amount of data through sample learning,
which has been independently labeled by analysts to generate the required output.
With the development of the system, deep neural network (DNN) has become the
latest technology in machine learning model. DNN can provide power for tasks such
as machine translation and speech or object recognition with higher accuracy. However, training DNN requires a lot of labeling data. The process of accumulating and
labeling a large amount of information is expensive and time-consuming. In addition
to the challenge of accumulating labeled data, most machine learning models are
fragile and vulnerable to collapse when their operating environments change slightly.
For example, in speech recognition systems, if the acoustic environment or the sensor
of the microphone changes, it may be necessary to retrain on a completely new
dataset. The time and effort required to adjust or modify the model are almost
equivalent to recreating the model.
Second is the problem of autonomous learning. At present, AI takes machine
learning technology as its core and relies on huge dataset. However, the battlefield
environment is complicated and changeable, the competition is violent, the situation is
changing rapidly, full of fog and accident, and the AI system cannot predict all possible
408
G. Zhou