Page 141 - Artificial Intelligence for the Internet of Everything

The Web of Smart Entities 127

eventually enable the automatic generation of diet data, we believe that there
will always be cases in which data will need to be entered manually. We
would like to point out that, in the case of video recognition, the data, while
technically coming from a sensor, requires sophisticated image processing.
Aggregated data. If we look at how “Google maps” ascertains traffic data, it
is simply the aggregate of data from cell phones in cars. There is certainly a
good amount of processing necessary to produce useful data about the
movement of phones in vehicles. Notice that “Google maps” uses this data
to eventually produce a model of congestion. However, before doing so,
“Google maps” does produce aggregate data.
Other models. We have seen several examples in which data from models
feed into other models and, as such, generate useful data for these other
models. For example, a model that is designed to balance fitness will need
access to the data from a model capturing diet data as well as a model cap-
turing exercise data. We imagine that a model that balances fitness would
furthermore interact with other models, such as calendars, vehicles, public
transportation and restaurants.
Aggregate models. Just as Google aggregates data from individual phones in
cars to construct a model of traffic flow, we can imagine cases in which we
wish to aggregate models. Consider models of exercise data. If we were
interested in simply ascertaining the overall exercise activities of a firm’s
employees, we would only need to gather a single data point from each
employee. However, if we wish to ascertain exercise patterns, perhaps in
the context of scheduling gym hours or to determine how big of a gym
to build, then models of exercise patterns are necessary.
Feedback loop. A feedback loop of a model to itself enables monitoring and
reflection on the workings of the model. Suppose a model of a person’s food
preferences is matched to someone else’s model. A recipe may be returned
that is deemed to match a person’s preferences. In case the person does not
like the recipe, or perhaps the matching parameters are insufficient or were
weighted improperly, we would like to adjust the model. We then think of
how case-based reasoning matches new cases to an existing case-base (see
Wikipedia, 2016).

7.5.2 Real-Time Models
A good number of smart devices already maintain real-time models. Con-
sider a Nest thermostat; it builds a model of a user’s heating and cooling pref-
erences. In particular it builds a real-time model as it constantly learns from