training, learn from past experience and adapt to the accumulated knowledge even in
the face of a new environment. From here we can see that even the most advanced AI
systems at this stage are still far from adaptive biological intelligence. In 2017, DARPA
launched a project called Lifelong Learning Machines (L2M) [8].
The learning mechanism proposed by L2M does not need to provide the system
with a large number of knowledge sets about applications as it does now. It needs
limited domain knowledge and methods to start its behavior and learning. L2M tries to
apply biological learning mechanism to machine learning system, breaking the
dependence of existing machine learning system on pre-programming and training
samples, so that artificial intelligence system, like biological system, can make decisions based on experience. Even in the new situation without experience and training
before, it can also make scientific decisions quickly according to “experience” accumulated in “life” and improve its operation. Active autonomy enhances the ability to
adapt to the environment. The L2M project will use the principles of computer science
and biological learning to build learning paradigms and evolutionary networks, and
continue learning through external data and internal goals. At the same time, L2M
project will set the limit of system capability and build a security mechanism for AI
system.
3.4 Setting Up Machine Commonsense Research to Enhance Reasoning
Ability of AI System
Common sense is a basic ability of human beings to perceive and understand the world.
Typical AI systems lack a general understanding of how the physical world works, a
basic understanding of human motivation and behavior, and a general understanding of
things like adults. The lack of common sense hinders the AI system’s understanding of
the world, its natural communication with humans, its rational behavior in unpredictable situations and its learning of new experiences.
At present, in all machine learning applications, machine reasoning is relatively
narrow and highly specialized. Researchers must train and program AI systems for each
situation, but commonsense reasoning is still missing. To this end, in 2018, DARPA
launched a project called Machine Common Sense (MCS) [9]. The MCS project will
explore the latest developments in cognitive understanding, natural language processing, in-depth learning and other areas of artificial intelligence research to find
answers to commonsense questions.
The MCS project plans to adopt two strategies to develop two different kinds of
commonsense services and to design a special evaluation method for the two services.
The technical area of research includes the following three aspects: First is the foundation of human common sense, learning experience to build computational models
that can simulate human perception of objects (intuitive physics), agents (intentional
actors) and locations (space navigation). These computational models will assess and
identify cognitive development milestones in developmental psychology research and
the literature to determine their learning effectiveness at three levels
(prediction/expectation, empirical learning and problem solving). The second is the
basic experimental environment of human common sense. The basic test environment
of human common sense will be used when testing the research results in the field of
Key Problems and Solutions of the Application …
411
the face of a new environment. From here we can see that even the most advanced AI
systems at this stage are still far from adaptive biological intelligence. In 2017, DARPA
launched a project called Lifelong Learning Machines (L2M) [8].
The learning mechanism proposed by L2M does not need to provide the system
with a large number of knowledge sets about applications as it does now. It needs
limited domain knowledge and methods to start its behavior and learning. L2M tries to
apply biological learning mechanism to machine learning system, breaking the
dependence of existing machine learning system on pre-programming and training
samples, so that artificial intelligence system, like biological system, can make decisions based on experience. Even in the new situation without experience and training
before, it can also make scientific decisions quickly according to “experience” accumulated in “life” and improve its operation. Active autonomy enhances the ability to
adapt to the environment. The L2M project will use the principles of computer science
and biological learning to build learning paradigms and evolutionary networks, and
continue learning through external data and internal goals. At the same time, L2M
project will set the limit of system capability and build a security mechanism for AI
system.
3.4 Setting Up Machine Commonsense Research to Enhance Reasoning
Ability of AI System
Common sense is a basic ability of human beings to perceive and understand the world.
Typical AI systems lack a general understanding of how the physical world works, a
basic understanding of human motivation and behavior, and a general understanding of
things like adults. The lack of common sense hinders the AI system’s understanding of
the world, its natural communication with humans, its rational behavior in unpredictable situations and its learning of new experiences.
At present, in all machine learning applications, machine reasoning is relatively
narrow and highly specialized. Researchers must train and program AI systems for each
situation, but commonsense reasoning is still missing. To this end, in 2018, DARPA
launched a project called Machine Common Sense (MCS) [9]. The MCS project will
explore the latest developments in cognitive understanding, natural language processing, in-depth learning and other areas of artificial intelligence research to find
answers to commonsense questions.
The MCS project plans to adopt two strategies to develop two different kinds of
commonsense services and to design a special evaluation method for the two services.
The technical area of research includes the following three aspects: First is the foundation of human common sense, learning experience to build computational models
that can simulate human perception of objects (intuitive physics), agents (intentional
actors) and locations (space navigation). These computational models will assess and
identify cognitive development milestones in developmental psychology research and
the literature to determine their learning effectiveness at three levels
(prediction/expectation, empirical learning and problem solving). The second is the
basic experimental environment of human common sense. The basic test environment
of human common sense will be used when testing the research results in the field of
Key Problems and Solutions of the Application …
411