Page 41 - Kaleidoscope Academic Conference Proceedings 2020

• The second is the sheer improvement in computing circumstances? It also begs the question of what is
power available for the ML training process (which I acceptable and what is not? If we were to perform a similar
won’t discuss here but for which there is plenty of task as human beings as the ML model is being asked to do
literature, an example being [13]). This step is we might not do any better, so where do we set the bar?
extremely compute-intensive, and computing in the
cloud has greatly democratized access to the required The last point I would like to bring up is that the main effort
resources. The “weights” determined by the training in using ML is not what goes into the ML model algorithms
process are subsequently used by an inference engine or into the computing. The overwhelming effort goes into the
to relate the pattern in an input to a specific identity, data
result, or action.
• The third is the inference engines (IE) that take
advantage of the asymmetry between the training of
an ML model and its execution. Such an inference
engine can run 4-6 orders of magnitude faster than the
time required for training. The IE can be purpose-
built hardware or software running on a general-
purpose computer. An IE can run many different
models and be used for many applications. This is the
device that uses the “weights”

What goes along with the algorithmic aspect of ML are the
necessary resources for computing, data storage, processes
used, operational aspects, and the time scales associated with
the resources. These are illustrated in Figure 3. Figure 4 - A ML learning network with hidden layers
and its life cycle. Typically, the fraction is 50% - 80% as
shown in Figure 5. The approach in dealing with training is
to keep the model as simple as possible. As an example, if
we want the model to recognize a good weld pool from
images in a welding process the training data may have a
simple annotation without explicit description of the weld
features; we leave that to the model to figure out. While I
suspect that adding such features as part of the training would
improve the results, such annotation is expensive to do and
just raises the data costs even further. If it were possible to
do such annotation automatically it might well be worth it;
this is an area where there is a lot of experimentation yet to
be done.
Figure 3 - Flows and processes for AI/ML applications

The allure of ML is that it does not require the hard work
needed by traditional methods and the specialized
knowledge that went into developing models or simulations
that are based on the basic laws of physics, chemistry, or
mechanics. As shown in Figure 4, what ML algorithms learn
is whether the input pattern is correctly identified with an
output. The training consists of input patterns used to set
weights within the network so that the outputs are clustered
under a correct label. The best success has been with images.
An example would be images of highway signs – so the input
could be pictures of “Stop Signs”, or “Yield Signs”, or
“Pedestrians Crossing Signs” taken under as many
conditions as possible. Success would be correctly assigning
the right label to each sign. There are, however, occasional Figure 5 - Typical allocation of effort from “Towards Data
surprises and inexplicably an ML model will produce an Science” - https://towardsdatascience.com/
unexpected and incorrect result, even though this happens
rarely. The root causes for such errors are in general not well
understood. This poses a natural question about setting
metrics and somehow determining error bounds to anticipate
what level of incorrect results is likely and under what

– xxxvii –