Page 70 - Artificial Intelligence for the Internet of Everything

56 Artificial Intelligence for the Internet of Everything

learning. The information in the samples is likely to be highly heterogeneous
in nature. Depending on circumstances, samples might consist of one or
more of the following data: still images in various parts of the spectrum
(IR, visible, etc.), video, audio, telemetry, solid models of the environment,
records of communications between agents, and so on. Multiple modalities
offer the potential to allow one type of data to help sort through ambiguities
and unobservability in another modality. Moreover, complex and wide-
spread events such as social movements can only truly be understood by
aligning and composing these heterogeneous sources to detect underlying
patterns (Giridhar, Wang, Abdelzaher, Al Amin, & Kaplan, 2017; Gui
et al., 2017).
The danger is that some samples may be misleading in general, even if
unintentionally (e.g., an action succeeds even though an unsuitable action
is applied) and the machine-learning algorithms will have to make the dis-
tinction between the relevant and irrelevant, the instructive and misleading.
In addition, some of the samples might be a product of intentional deception
by the enemy. In general, issues of adversarial learning and adversarial rea-
soning are of great importance (Papernot et al., 2016). Although initial dem-
onstrations that showed how to perturb an image to fool a trained model
were seen as an amusement, later demonstrations have shown that not only
can these attacks be used in the real environment (Kurakin, Goodfellow, &
Bengio, 2016), they also do not require insider knowledge of the learning
system (Papernot et al., 2017) and are exceedingly difficult to detect in iso-
lation (Carlini & Wagner, 2017). Robustness to these adversarial actions will
come about as the product of engineering principles that treat learning
as part of a larger system of models and heterogeneous data that provide
avenues to check and attest to the veracity of our models.

3.5 AI ENABLES EMBODIED AGENTS
Some intelligent things will be embodied so that they can actively explore
and interact with the world, not merely constructs that protect our virtual
environment or sensors that vigilantly observe and interpret. AI enables
these interactions, allowing things to develop intuition about the laws of
physics and to learn how to act optimally to accomplish their missions. In
some cases, agents will take our traditional hand-crafted control paradigms
and expand them through learning and artificial evolution; in other cases,
agents will learn behaviors as a child does, developing motor skills through