The Dos and Don'ts of Machine Learning Deployment: A Key Strategic Business Issue
ML is one of the keys for the future success of any business
ML is a transformative technology which has the potential to impact businesses in every dimension, and its adoption is rapidly becoming critical for the competitiveness of enterprises and industries, if not for their survival. This is because of the role that ML can have in delivering better customer experience and improving productivity.
ML does not concern only high-tech companies, IT companies or companies with huge R&D budgets like Google and Microsoft. Every organization should consider how ML fits to its strategy and operations, and investigate how this technology can be applied to the organization’s business.
No time to lose
When to start the ML journey? My answer is simple: don’t wait! First, because there are significant opportunities to seize. Second, because successful ML deployment takes time, and needs trial and error. It can take weeks and most likely months before a deep learning model is trained. The good news is that it is not too late: a worldwide survey of 196 organizations by Gartner (2018) on advances in data and analytics showed that 91 percent of organizations have not yet reached a ‘transformational’ level of maturity, despite this area being a number one investment priority for CIOs in recent years. So, if your organisation is not surfing the wave yet, you’re not alone.
The strategic imperative
Acquiring ML tech is indeed not enough. Its successful application is an organisational and managerial transformation challenge.
How will we take our organisations and our industry through the ML journey? No application of ML can have guaranteed success. However, there are five general strategic guidelines that I believe are worth bearing in mind to minimise the risks of failure and build the strong and sustainable foundations.
- Ask the right business question that ML can help with in the short term
Examples of common mistakes endangering the successful start of the ML journey include developing solutions that add limited value to the business, generating insight that it is too risky or difficult to act upon, and failing to make sufficient tangible and valuable progress in the first one to two years.
To avoid these, it is essential to think about business priorities and where ML strengths are, for example classification type problems. “Reducing costs” or “increasing customer satisfaction “may be too broad objectives though. It is essential to frame the problem or opportunity into something specific and something that data, and ML applications in particular, can help solve. For instance, asking how ML can help better predict and mitigate disruptions to train services was the starting point that RSSB took in the context of the “Data Sandbox” competition.
- Sort out ongoing access to new training data
In order to successfully apply ML, rich and quality data is essential, and this is no different from other data analysis and modelling approaches, but it is not all. ML models can quickly become dated (even properly trained ones) and are bound to give the wrong answers if not refreshed with new training data that reflect changes in people’s preferences, maintenance practices, operational approaches, etc. Businesses that succeed in deploying ML applications manage to establish a ‘virtuous circle of data’: ML algorithms are trained, tested, put to use, and then continuously retrained and improved with fresh user data … and so forth.
This means that access to lots of initial good data is not enough. There is a need to keep data flowing into the ML model, and any ML application should consider this point. And when it is not obvious how fresh good training data will be generated and fed back, special consideration should be given on how to create some feed-back loops to ensure that the results remain valid while the world changes.
- Explore all options to get and retain the right talent, including upskilling
We have covered the skills shortage issue in a previous article. I would want to emphasise that even if the recruitment / partnership of ML experts is essential, not everything can be outsourced. How to get and keep the right talent will depend on the ability of our organisations to accompany current staff on the upskilling and data literacy journey, for instance by using some of the dedicated ML platforms that have mushroomed in the past years.
ML use should not be confined to technical or IT teams, ML is for the whole business. Blake Morgan, Customer experience futurist, explains how Amazon’s flywheel approach to AI and ML means that innovation in one area of the company fuels the efforts of other teams. Essentially, what is created in one part of Amazon acts as a catalyst for other areas and teams to drive their product innovation and create a cohesive customer experience.
- ML successes are all about not overdoing things, and being a ‘pragmatic strategist’
It is essential to build ML capability little by little, step by step. One has to enter into the ML journey on a -proof-of-concept basis, and be strategic in what is chosen - only a few applications will be game changing and deliver quick win in sufficiently short time scales to fuel energy and funds for the next steps of the journey.
And, there is the need to build experience. Nothing will replace this.
- Understanding and embracing ML at the top
Last but not least, the strategy should be introduced and guided at the C-suite level, and I would even say, at the CEO level. For ML to be adopted wisely and successfully, its basis should be understood by the people responsible for the business strategy. In the rail industry, CEOs most of the time come from a rail operational or engineering background. The business and the science of data are often associated with IT systems and heavily delegated.
In our industry it is fully recognised that an engineering team cannot succeed without operational knowledge; and similarly, there is very little questioning that senior operation specialists would not be able to make the right decisions without the awareness of engineering capabilities and constraints. But what about the following question: will engineers and operators be able to deliver a better railway in the future without a sound grasp of what data science, and ML techniques in particular, can do for them? Given what we’ve seen throughout this blog series, it seems very doubtful. Rail needs to make sure that action is taken now as skills and experience take time to build.
