Instrumentation & Measurement Magazine 24-2 - 88

It's All in the
Definition
With the above in mind,
Table 2 gives some I&M
and ML terminologies and
their relation when ML is
used in a measurement system. Experts will notice
some conceptual simplifiFig. 5. The concepts of under- and overfitting the model to the dataset. All three parts have the same dataset, indicated
by the black dots, and are trying to fit a model to that data, indicated by the blue curve.
cations for both I&M and
ML. This was done on purand Logistic Regression are known to have low variance but pose to make each more understandable to the experts of the
high bias.
other's community. A warning is in order here: although we
It is important to understand why ML has bias and variance have used the most-commonly used definitions available in
in the first place. In ML, bias occurs because either the model is the ML literature, the fact is that ML does not have any internot a good fit for the dataset, as shown by the low-complexity national standards and guidelines for its terminology, unlike
model (linear fit) in Fig. 5a, or the training dataset is not a good I&M which has standards and guidelines such as [4], [11], [14],
representation of the general data. In ML terms, we say that in [15] and others. In fact, we came across a few different and
this case the model underfits the data. The more the model un- even contradictory definitions for some ML terminology in
derfits, the higher its bias because, as we can see in Fig. 5a, the the ML literature. As such, readers will also likely come across
differences between many of true values and their correspond- some terminology interpretations that are different from what
ing estimated values become quite large which, according to is given in Table 2 for ML.
(2), leads to a large bias. On the other hand, variance occurs
if we fit the model to the training data (and consequently its Measurement Result Prediction:
embedded noise) too closely, as shown in the high-complex- Measuring without Measurement
ity model (high-degree polynomial fit) in Fig. 5c, causing wild It should be mentioned that it is possible to use ML alone, withfluctuations in the output. In ML, we say that in this case the out the direct measurement component shown in Fig. 1, to
model overfits the data. The more the model overfits, the higher predict a measurement result. One example is [21], which takes
its variance because, as we can imagine from Fig. 5c, the dif- as input the IP addresses of two nodes on the Internet and inference between the estimated value for input x in one dataset stantly predicts the end-to-end network delay between those
compared to the average of all estimates in all datasets for the two nodes. In this system, the ML model has been trained with
same input x becomes large (compared to Fig. 5a) due to the a large dataset consisting of end-to-end delays directly meawild fluctuations, and this large difference leads to a large vari- sured between a huge number of IP addresses on the Internet.
However, in this and similar ML-alone systems, no experiance according to (3).
Fig. 5 is a representation of regression ML methods, which ment occurs at run time, and because a purely mathematical
fit a continuous curve to the
training data, and for a given
input give its corresponding
output value on the curve.
There is another class of ML
methods called classification,
which groups the training
data into different classes,
as shown in Fig. 6, and for a
given input estimates the class
where that input falls into.
In classification, training the
model will lead to the decision boundary shown in the
figure. In I&M, ML classification is used for measuring
ordinal quantities, for example, measuring the quality of
a product from 1 (lowest quality) to 4 (highest quality), as
Fig. 6. An ML classification example. The dots come from a training dataset that indicates a product's quality level (1
to 4) based on two quality assurance tests performed during the manufacturing process.
shown in Fig. 6.
88	

IEEE Instrumentation & Measurement Magazine	

April 2021
Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2
