Page 68 - Artificial Intelligence for the Internet of Everything

54 Artificial Intelligence for the Internet of Everything

Fig. 3.3 AI-enabled agents—members of a human-agent team—will rapidly learn in
ever-changing, complex environments, providing the team’s commander with real-
time estimates of enemy, reasoning on possible courses of action, and tactically
sensible decision.

extreme physical and cyber threats. They must be effective in this unstruc-
tured, unstable, rapidly changing, chaotic, and rubble-filled adversarial envi-
ronment; learning in real-time under extreme time constraints, with only a
few observations that are potentially erroneous and with uncertain accuracy
and meaning, or are even intentionally misleading and deceptive (Fig. 3.3).
It is clearly far beyond the current state of AI to operate intelligently in
such an environment and with such demands. In particular machine learn-
ing, an area that has seen dramatic progress in the last decade, must experi-
ence major advances in order to become relevant to the real battlefield. Let’s
review some of the required advances.
Learning with a very small number of samples is clearly a necessity in an
environment where friends and enemies continuously change their tactics
and the environment itself is highly fluid, rich with details, dynamic, and
changing rapidly. Furthermore, very few (if any) of the available samples will
be labeled or, if so, not in a very helpful manner. This learning stands in stark
contrast to the highly influential ImageNet dataset (Deng et al., 2009) that
led to the advent of modern deep learning by providing a rich, labeled