Page 11 - Machine Learning for Subsurface Characterization

Preface xxiii

capabilities of their products, there are three common customer responses: (1)
product capabilities are overhyped, (2) vendors do not know what they are talk-
ing about, and (3) customer is exhausted with yet another machine learning
product. There is a perception that, when vendors are promoting their “AI plat-
forms,” these are repackaged versions of traditional business intelligence or
analytical tools. The big data hype a decade back was very similarto the current
AI/ML wave. That time, big data was the new business intelligence, and big
data was projected to solve everything [2].
Another big challenge to the adoption of ML is whether the technical
domain experts will trust and adopt the ML models. Domain experts have a lim-
ited understanding of the reasons due to which an ML system makes a certain
recommendation, in other words explain how an answer or insight was pro-
duced by the ML system. From a domain expert’s point of view, the ML pre-
dictions/results are too generic and nonspecific—lacking deep insights.
Without the interpretability of the model and explainability of the results, there
is some level of faith that must be put in ML-based strategies. The biggest bar-
rier to machine learning in several industries is the culture that values a domain
expert’s intuition over data-driven solutions. Quite often, the problem lies in
how AI, ML, and DL are portrayed, for example, when it is said “neural net-
works are inspired by neurons in the brain” or “convolutional neural networks
are inspired by human visual processing system.” It is not clear to anyone out-
side of the inner circle how to start applying the AI technologies. They should be
made aware that the barriers to entry are quite low. AI, machine learning, and
deep learning are not hard-to-grasp, science-fiction concepts but are based on
mathematical, statistical, and computational foundations. ML and DL may be
hard to understand but are indeed simple to implement (caution: implementa-
tion is easy, but the evaluation is very challenging).
ML methods work well when a complex task requiring human intelligence is
broken into simpler less-intelligent, pattern recognition–type problems. For
example, machine learning can fill missing words in a sentence and translate
the sentences to different languages; however, these methods are far from deriv-
ing the concept or intent of a sentence. ML methods tend to be effective for nar-
rowly focused tasks. For example, ML-assisted conversational chatbots can
now perform goal-oriented conversations: setting up an appointment between
two people, wherein the goal is limited to coordinating the calendars of two
people.
Irrespective of successes and failures, efforts to infuse ML into organiza-
tions are spreading like wildfire and are a reality. ML implementations, design,
and approaches are evolving too fast; the ML practitioners are having trouble
staying abreast of leading ML practices. In a haste primarily driven by the fear
of missing out, organizations are entering the ML race without sufficient plan-
ning. A premature adoption of ML adversely affects the outcomes, lending a
bad name to machine learning. For avoiding such scenarios, Harvard Business
Review (HBR) suggests taking a portfolio approach to truly harness machine