3.6 Revealing the Transparency of AI Decision-Making and Enhancing
Credibility by Improving the Interpretability of AI System
Interpretability of AI systems has become a key obstacle to human–computer cooperation. Explainable Artificial Intelligence (especially explainable machine learning)
will be indispensable if we want human beings to understand, trust and effectively
manage the new generation of AI systems. To this end, in October 2016, DARPA
launched the Explainable Artificial Intelligence (XAI) project [12]. The goal is to
establish new or improved machine learning technology, generate explainable models
and integrate effective interpretation technology to enable end users to understand, trust
and effectively manage future AI systems. Through this project, the new machine
learning system will be able to explain its own logic principles, describe its own
advantages and disadvantages, and explain future behavior.
Explainable AI faces three challenges: how to generate explainable models, how to
design interpretive interfaces and how to understand users’ psychological needs for
effective interpretation.
In response to the first challenge, the project will develop a series of new or
improved machine learning technologies to generate more interpretable models; in
response to the second challenge, it hopes to integrate the latest human–computer
interaction technologies (such as visualization, language understanding, language
generation and session management) with new principles, strategies and technologies
for effective interpretation; in response to the third challenge, the project will plan to
summarize, expand and apply current psychological theories of interpretation.
At present, the project has achieved some results. At the IJCAI/ECAI 2018
Workshop on Explainable Artificial Intelligence (XAI) conference held in July 2018,
several papers funded by the project were published, including depth adaptive programming through reward decomposition interpretation, natural language interpretation
of visual question and answer using scene diagrams and visual attention generation,
finite state representation of learning cycle strategy network, etc.
4 Conclusion
Artificial intelligence technology has shown good application prospects in all walks of
life and has become the core and key to achieve leap-forward development of every
industry. Military field is no exception. According to the characteristics of military
field, four key problems of AI application in military are analyzed and summarized,
including data problem, autonomous learning problem, security problem and credibility
problem. For each of these problems, through the analysis and summary of domestic
and foreign research, the corresponding solutions are introduced, which can provide
reference to related research.
Key Problems and Solutions of the Application …
413
Credibility by Improving the Interpretability of AI System
Interpretability of AI systems has become a key obstacle to human–computer cooperation. Explainable Artificial Intelligence (especially explainable machine learning)
will be indispensable if we want human beings to understand, trust and effectively
manage the new generation of AI systems. To this end, in October 2016, DARPA
launched the Explainable Artificial Intelligence (XAI) project [12]. The goal is to
establish new or improved machine learning technology, generate explainable models
and integrate effective interpretation technology to enable end users to understand, trust
and effectively manage future AI systems. Through this project, the new machine
learning system will be able to explain its own logic principles, describe its own
advantages and disadvantages, and explain future behavior.
Explainable AI faces three challenges: how to generate explainable models, how to
design interpretive interfaces and how to understand users’ psychological needs for
effective interpretation.
In response to the first challenge, the project will develop a series of new or
improved machine learning technologies to generate more interpretable models; in
response to the second challenge, it hopes to integrate the latest human–computer
interaction technologies (such as visualization, language understanding, language
generation and session management) with new principles, strategies and technologies
for effective interpretation; in response to the third challenge, the project will plan to
summarize, expand and apply current psychological theories of interpretation.
At present, the project has achieved some results. At the IJCAI/ECAI 2018
Workshop on Explainable Artificial Intelligence (XAI) conference held in July 2018,
several papers funded by the project were published, including depth adaptive programming through reward decomposition interpretation, natural language interpretation
of visual question and answer using scene diagrams and visual attention generation,
finite state representation of learning cycle strategy network, etc.
4 Conclusion
Artificial intelligence technology has shown good application prospects in all walks of
life and has become the core and key to achieve leap-forward development of every
industry. Military field is no exception. According to the characteristics of military
field, four key problems of AI application in military are analyzed and summarized,
including data problem, autonomous learning problem, security problem and credibility
problem. For each of these problems, through the analysis and summary of domestic
and foreign research, the corresponding solutions are introduced, which can provide
reference to related research.
Key Problems and Solutions of the Application …
413