captures data, compares the user’s use of cover and range crossing areas, then evaluates
the user’s performance based on other data and records these data into a database on the
combat methods of US military professionals. Finally, Athena evaluates the data and
provides constructive suggestions for users, comparing their performance with that of
top-level users.
Through this game environment, the USA seeks competent commanders and
obtains the data needed to build future AI applications. Once enough data is available,
Athena can simulate modern military operations and propose new tactics.
Athena provides a test platform for exploring how to integrate AI into military
decision-making process. Using war games to observe how military experts make
decisions, benchmark data is established for testing a series of applications that enhance
warfighter judgment.
3.2 Explore a New Learning Algorithm to Reduce the Dataset Needed
for Model Training and Solve the Problem of Data Annotation
In order to reduce the cost and time associated with training and adjusting machine
learning models, DARPA launched the project of Learning with Less Labels (LwLL)
[7]. Through the LwLL project, DARPA will study new learning algorithms to greatly
reduce the dataset for model training or updating. This project aims to make the process
of training machine learning model more efficient. DARPA said that through the LwLL
project, the dataset needed to build a model will be reduced by six orders of magnitude
or more, and the label precedents used to label the model will be adjusted from millions
to hundreds; that is, a million images are needed to train a system now, and only one
image or about 100 label precedents will be needed in the future.
To achieve this goal, the LwLL project will explore two technical areas. The first
technology area focuses on building efficient learning and adaptive learning algorithms.
The main approaches are innovation in meta-learning, transfer learning, active learning,
K-shot learning and supervised/unsupervised adaptation. The second technical area
explores the normative description of machine learning problems, including the difficulty of decision-making and the real complexity of the data used to make decisions. In
this regard, DARPA said that it is difficult to assess the efficiency of building machine
learning systems or the basic limitations of the accuracy level of the model. Through
LwLL project, we hope to find the theoretical limit of machine learning possibilities
and use this theory to promote system development and enhance capabilities.
3.3 Enhancing AI System’s Autonomous Learning and Environmental
Adaptability by Bionics
At present, AI takes machine learning technology as its core and relies on huge data
support. But the real world is full of contingency, and programmers cannot predict all
possible factors or situations. When these machine learning systems encounter special
situations that are not included in programs and databases, they will be at a loss. If we
want to expand the capability of machine learning system in this new environment, we
must stop its service and retrain it with additional data, but this method relies on human
intervention and is inefficient. In contrast, biological systems can conduct independent
410
G. Zhou
the user’s performance based on other data and records these data into a database on the
combat methods of US military professionals. Finally, Athena evaluates the data and
provides constructive suggestions for users, comparing their performance with that of
top-level users.
Through this game environment, the USA seeks competent commanders and
obtains the data needed to build future AI applications. Once enough data is available,
Athena can simulate modern military operations and propose new tactics.
Athena provides a test platform for exploring how to integrate AI into military
decision-making process. Using war games to observe how military experts make
decisions, benchmark data is established for testing a series of applications that enhance
warfighter judgment.
3.2 Explore a New Learning Algorithm to Reduce the Dataset Needed
for Model Training and Solve the Problem of Data Annotation
In order to reduce the cost and time associated with training and adjusting machine
learning models, DARPA launched the project of Learning with Less Labels (LwLL)
[7]. Through the LwLL project, DARPA will study new learning algorithms to greatly
reduce the dataset for model training or updating. This project aims to make the process
of training machine learning model more efficient. DARPA said that through the LwLL
project, the dataset needed to build a model will be reduced by six orders of magnitude
or more, and the label precedents used to label the model will be adjusted from millions
to hundreds; that is, a million images are needed to train a system now, and only one
image or about 100 label precedents will be needed in the future.
To achieve this goal, the LwLL project will explore two technical areas. The first
technology area focuses on building efficient learning and adaptive learning algorithms.
The main approaches are innovation in meta-learning, transfer learning, active learning,
K-shot learning and supervised/unsupervised adaptation. The second technical area
explores the normative description of machine learning problems, including the difficulty of decision-making and the real complexity of the data used to make decisions. In
this regard, DARPA said that it is difficult to assess the efficiency of building machine
learning systems or the basic limitations of the accuracy level of the model. Through
LwLL project, we hope to find the theoretical limit of machine learning possibilities
and use this theory to promote system development and enhance capabilities.
3.3 Enhancing AI System’s Autonomous Learning and Environmental
Adaptability by Bionics
At present, AI takes machine learning technology as its core and relies on huge data
support. But the real world is full of contingency, and programmers cannot predict all
possible factors or situations. When these machine learning systems encounter special
situations that are not included in programs and databases, they will be at a loss. If we
want to expand the capability of machine learning system in this new environment, we
must stop its service and retrain it with additional data, but this method relies on human
intervention and is inefficient. In contrast, biological systems can conduct independent
410
G. Zhou