Page 260 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

252 CHAPTER 12 Computational Intelligence in the Time

Passive solutions do not suffer from false positives and negatives in detecting a
change in stationarity/time variance, since the learning-based system/application
continuously updates parameters over time by integrating the available novelty
content into the model. From this perspective the partial derivative in Eq. (12.2)
can be intended as the operator extracting the novelty information content from
current data: if there is novelty, the model needs to be updated. As a consequence,
passive models are rather sensitive to noise in incoming data: after parameter
convergence, the application continues to update the parameters so as to track the
particular noise realization. A vast literature on passive approaches exist (also called
online learning in the neural networks field) and the interested reader can focus on
and implement those results in her/his CPS/IoT application [4,11,14]. However, the
computational complexity of a passive approach might be inappropriate for
embedded applications given the continuous need to update the model; operations
which might end up in a prohibitive energy consumption and processing time
whenever the update phase is energy/computation eager. Ensemble solutions and
complex neural networks are examples in this direction where the cost we have to
pay for high accuracy and flexibility is computational.

4.2 ACTIVE ADAPTATION MODALITY
In the active adaptation strategy, the presence of a change-detection trigger activates
the due application reaction following the detected change in stationarity, for
example, by updating the learning-based system. This means that the application
running on the embedded device undergoes an update/reconfiguration phase to track
the change in stationarity only when triggered by the change detection module. The
change detection moduledor Oracledoperates by inspecting features 4 extracted
from the input data or preprocessed variables [4,13] to assess the presence of a
change. In other terms the Oracle u acts as the indicator function

1; if change in stationarity is detected
uð4Þ¼
0; otherwise
and the update equation becomes

vLðx; qÞ
q tþ1 ¼ q t guð4Þ (12.3)
vq
q t
In its basic form the Oracle is based on a threshold; in more sophisticated versions,
the Oracle also relies on a confidence level, for example, as it happens in statistical
testsdor, even, takes control of the occurrence of false positives by setting its
expectation to a predefined value.
Adaptive strategies at CPSs following the active adaptation modality are known
as “detect & react” approaches; such solutions have much focused on classifiers in
the related literature, though results are more general [4].
It should be commented that if the computational load of passive solutions is
negligible they should be preferred than active ones unless the application is