Perception: Machine learning is more hype than reality.
Reality: In the Gartner Inc. hype cycle graph, machine learning (ML) currently features in the “peak of inflated expectations” phase. This phase comes with a prediction that ML will become an integral part of business processes across industries in finance, HR, supply chain, sales, marketing, and service.
As a classic case of a self-fulfilling prophecy, this prediction has led to a change in the behavior and actions of many organizations as they reimagine business processes for greater intelligence and automation. Even the organizations that have not yet fully explored the power of basic analytics are now conceptualizing and piloting ML projects in the same way mature organizations are.
For that reason, what we see now is a mix of successes and failures. There are organizations that have successfully deployed ML to create valuable customer oriented solutions and improve their business processes. And there are organizations that have failed to deliver on ML projects. However, irrespective of successes and failures, efforts to infuse ML into organizations are spreading like wildfire and are a reality.
In such a scenario, it is possible that organizations enter the ML race without undergoing sufficient coaching and conditioning. This premature adoption of ML may affect outcomes, resulting in sporadic dehydration. To address this challenge, Harvard Business Review (HBR) suggests taking a portfolio approach to artificial intelligence (AI) and ML projects. As per HBR, there should be a mix of projects, ones that have the potential to generate quick wins and long-term projects for end-to-end transformation of business processes. With such an approach, there is a greater chance of more ML efforts being realized.
Perception: Hiring the best ML talent is sufficient to build great ML solutions.
Reality: Certainly, data science or ML is the most sought-after skill set in the market today. A great ML practitioner who understands the basic pitfalls of an analyst’s life, such as correlation-causation or biases, and who has sufficient business understanding could bring significantly higher value to business solutions. Consequently, organizations are fast building data science teams for ML adoption.
But setting up best teams alone is not sufficient. An equally important yet obvious fact that is often missed is the phenomenon of relevant data, both in quality and quantity. Google LLC, Facebook Inc., Netflix Inc., and Amazon.com Inc. are powerful not just because of their intelligent algorithms, but also because of the data they have about people interacting with them.
In my experience of working with companies, I have seen data science teams waiting for long periods of time for the right data to build models for various use cases. To cite an example, one team in a bank spent a huge amount of time hammering on the available data points to build a model to understand the root causes of stagnation of national electronic funds transfer (NEFT) transactions and the rise of check-based transactions. None of the drivers suggested by the model appeared to be a root cause and, therefore, did not make sense to the team. As a next step, the team went about personally interviewing customers. The finding was interesting – in check-based transactions, people earn interest on their savings account balance for the three to four days it takes for the check to clear, compared to no interest in NEFT-based transactions. This data point was missing from the initial set.
This is a very simple example of how data is equally, if not more, relevant to building good ML models than skill sets alone. In fact, to highlight the importance of data in building great ML solutions, John von Neumann, a computer scientist, made a mocking comment, “With four parameters, I can fit an elephant, with five, I can make him wiggle his trunk.” His reference is, of course, to people building complex models with too few data points. The result is a picture no more accurate than a child’s attempt at drawing animal patterns in a starry night sky (as per Nate Silver’s book, Signal and Noise).
Perception: ML deployment requires huge computing power and skill sets.
Reality: The key building blocks of any successful ML deployment include the availability of huge computing power, skill sets, and data. A few years ago, the procurement of these building blocks was a challenge in terms of cost, serving as a significant barrier to ML deployment. The situation is markedly different today with technology vendors coming up with ready-to-deploy ML offerings and infrastructure in the cloud.
Today, organizations can leverage the computing power of the graphics processing units (GPUs) required for ML tasks through a subscription model. Through a microservices framework, one can simply call for pretrained ML models for image classification, topic detection, or time-series change-point detection over simple Web APIs that are ready for immediate use. Lastly, there are ready-to-use ML business applications to make enterprise customer service, finance, HR, and marketing applications more intelligent.
To summarize, ML today is undoubtedly delivering on its hype by creating innovative and valuable solutions for organizations. There are enough success stories out there to keep up the ML efforts. However, failures also abound, from which companies can learn to improve their traction.
For more on this topic, see Machine Intelligence Ascending, Part 1.