Page 286 - Artificial Intelligence for the Internet of Everything

264 Artificial Intelligence for the Internet of Everything

stated goal within specified constraints. The size and diversity of the IoT
ecosystem will ensure that humans cannot easily design such systems, espe-
cially given that many systems will be designed for one-off uses with
on-the-fly provisioning.
This means that we will require, at a minimum, substantial artificial assis-
tance in IoT system design. Whereas before, the state space for a learning
problem could be extracted from a given architecture, now the search is over
the architectures themselves. In this chapter we discussed the logical seman-
tics of contracts; one can also interpret operads in terms of dynamical and
stochastic systems, allowing us to build rich semantic representations of
component capabilities into the learning problem.
Another point of contact concerns the multiplicity of perspectives dis-
cussed in Section 13.2. AI systems may interact, especially across scales, so
that the parameters of one learning problem may be determined as the out-
put of another. These interactions go both ways, with top-down modifica-
tions to operating parameters at lower levels (e.g., peak use incentives for
electricity consumption) and bottom-up prediction for aggregate systems
(e.g., monitoring expected consumption).
We can also use architectures to document and manage our applications
of AI. Any learning algorithm is itself a system, and we can use compositional
architectures to describe it. Just as they did for more general systems, these
provide a generic representation of algorithms, which is nonetheless flexible
enough to specialize for any particular learning problem. Furthermore,
recent research indicates that some AI algorithms are themselves composi-
tional, giving a functorial analysis of the method of backpropagation (Fong
et al., 2017). This could allow for more modular AI architectures, in which
certain components of an AI system can be swapped in and out without
relearning the entire model.
By clarifying the interactions between learning agents, humans, and
other system processes, this documentation can be especially useful
when we want to integrate learning processes and learned models into
larger systems that already exist, especially those involving joint cogni-
tive systems. Fig. 13.3 shows several examples of compositional process
architectures, in which a process is decomposed into subprocesses. In
this context, the boundaries between components can be regarded as
input/output resources passed between the subprocesses. As before,
we represent physical resources with solid lines and informational
resources (data) by dashed line, noting that the latter can be copied
whereas the former cannot.