Page 72 - Artificial Intelligence for the Internet of Everything

58 Artificial Intelligence for the Internet of Everything

and learn to generalize (Hefny, Downey, & Gordon, 2015). These models
then become an intimate part of the planning and control process (Boots,
Siddiqi, & Gordon, 2011; Williams et al., 2017).
Traditional scientific and design methodology emphasizes the need to
model and analyze the parts of a system in isolation, carefully controlling
the variables so that the most concise and elegant description of a phenom-
enon or capability can be developed. In the future where components are
the result of learning processes that can radically reshape their functionality,
these traditional processes must also evolve. This evolution is already taking
place with the advent of so-called end-to-end learning approaches that build
and simultaneously train the entire processing and decision-making pipeline
from perception to action (Bojarski et al., 2016; Levine, Finn, Darrell, &
Abbeel, 2016). Simultaneously learning on many tasks at once keeps this
process from overspecializing (Devin, Gupta, Darrell, Abbeel, & Levine,
2017)—another example of meta-learning. Eventually these techniques will
be able to recapture the crucial feature of traditional systems engineering:
being able to formally prove correctness (Aswani, Gonzalez, Sastry, &
Tomlin, 2013).
Despite the future advances in embodied intelligence just discussed, the
challenge of limited electrical power will remain a driving factor in deploy-
ment for the battlefield. Most successful AI relies on vast computing and
electrical power resources including cloud-computing reach-back when
necessary. The battlefield AI, on the other hand, must operate within the
constraints of edge devices, such as small sensors, micro-robots, and the
handheld radios of warfighters. This situation means that computer proces-
sors must be relatively light and small, and as frugal as possible in the use of
electrical power. One might suggest that a way to overcome such limitations
on computing resources available directly on the battlefield is to offload the
computations via wireless communications to a powerful computing
resource located outside of the battlefield (Kumar & Lu, 2010). Unfortu-
nately, it is not a viable solution because the enemy’s inevitable interference
with friendly networks will limit the opportunities for use of reach-back
computational resources. We must turn to techniques to trade precision
for power to keep computing at the edge (Gupta, Mohapatra, Park, Raghu-
nathan, & Roy, 2011; Iandola et al., 2016) or encode mature AI procedures
directly into circuitry. Perhaps the systems will again lead the way, just as
they did when they used meta-learning to learn adaptability by obtaining
the correct balance between power and accuracy and tuning their own algo-
rithms to realize it.