Page 12 - Machine Learning for Subsurface Characterization

xxiv Preface

learning. HBR recommends having a mix of projects, ones that have the poten-
tial to generate quick wins and long-term projects for end-to-end transformation
of business processes. ML models perform the best after being exposed to large
historical/real-time datasets and stringent evaluation for a certain duration of
time, which ensures that the ML models are robust to edge cases and that they
do not pick up inconsistent patterns.

Recommendations for harnessing the power
of machine learning
1. When you master machine learning techniques, you can truly benefit from
the ever-growing vast datasets available to you. ML is a great tool to have at
your disposal, like computers, word processors, and mobile phones. As
computing speeds are expected to double five times over the next 10 years,
machine learning tools will serve as inexpensive tool to extract information
and insights from the enormous troves of data.
2. When you plan on using ML techniques, ensure you have a large, high-
quality dataset both to build the data-driven models and to test them. Also,
you need to ensure that the dataset you are using for building the models
should be available in the real world for ensuring a robust deployment of
the ML models. It could be argued that the data is more important than
the ML algorithms because ML algorithms are only as good as the data that
go into them. For example, Google, Facebook, Netflix, and Amazon are
leaders in ML applications not only because of their intelligent algorithms
and skilled data scientists but also because of the high-quality digital data
they have about people and products.
3. A vendor’s demo of ML workflow may work well on the vendor’s data; this
does not mean that the vendor’s ML workflow will work equally well when
applied to your data. Even when you see great results with your data, the
real-world deployment of the ML workflow will unearth severe limitations
in the ML implementations. Nonetheless, your efforts to fix these challenges
will make your ML implementations more robust.
4. Domain knowledge is a very important ingredient in building effective ML
models. ML users should know the limitations of ML methods and when
these methods can go wrong, or else, ML methods will learn relationships
that are totally spurious or tend to get overtrained without us knowing about
such gross errors. ML users should be aware that ML methods can pick up
patterns and relationships that are inconsistent without any physical basis.
5. ML tools are very good at learning clearly defined tasks, like identifying
people in photographs or accurately transcribing speech. ML tools currently
cannot understand human motivations or draw nuanced conclusions. For
now, ML methods work well when a complex task requiring human intel-
ligence is broken into simpler less-intelligent, pattern recognition–type
problems.