How To Properly Integrate Machine Learning Into Your Enterprise

'The assumption that you can just throw money at infrastructure is a dangerous one.' Sashi Marella Ph.D., senior data scientist at Viacom.


Despite the popularity and importance placed on machine learning and AI in enterprise, very few companies are successfully implementing these systems. Sashi Marella, the senior data scientist at Viacom, addressed the issue in his presentation at the Chief Data Officer Summit in New York.

He began by pointing out one of the biggest misconceptions people have when it comes to machine learning, noting that, 'There's an implicit assumption that somehow, you can throw more money you throw at the infrastructure data and it will provide us what we need and ROI's.'

The biggest problem with this assumption is that it immediately takes the onus to do more away from the stakeholders as they then consider their data science strategy completed. However, this is far from the case and its flaw is perfectly encapsulated by Ali Ghodsi of UC Berkeley, who said, 'there are only 1% that are succeeding in AI, the rest of the 99% are left behind and struggling to get all this big data technology.' The truth is, most companies have not taken the necessary steps needed to activate machine learning.

In order to begin to scratch the surface, you need to start empowering the data scientists within your organization in a meaningful way. Both data scientists and stakeholders, while more than proficient in their respective roles, won't be able to maximize their collective potential without a deeper insight into what the other team does.

At Viacom, Marella and his team are seven years into a program to develop their own in-house data platform. Through the course of it, he has come across two core problems common to any enterprise:

1. Price - executives and stakeholders have a tendency to go for the cheapest options available. Whether it be infrastructure or the implementation of it, the priority is whats cheapest.

2. Steep learning curve - If the company in question isn't tech-based or from a Silicon Valley background, the enterprise tends to difficulty wrapping their heads around machine learning, so much so, many are hesitant to even start.

These ' blockers', as Marella calls them, greatly hinder a company's ability to take advantage of AI-based solutions. For an AI-strategy to be effective, stakeholders have to relinquish some of their decision making power to algorithms they likely don't fully understand. Without this understanding of how machine learning is arriving at its solutions, it makes it impossible for them to even accurately weigh the risks of not going with the insights yielded.

The question then becomes 'how do we navigate this and create a successful data strategy which takes into account not only infrastructure but also a long-term view of the enterprise's goals.'

At Viacom, they ended up grouping these goals into 2 categories:

1. Autonomy of decisions - How do we educate stakeholders on the potential of machine learning so they are confident enough in their understanding to allow it to make determinations on behalf of the company.

2. Healthy trust - To make sure that it is a place of healthy understanding and not blind faith. This means an awareness of the risks involved with going with machine learning versus deciding not to.

Marella is quick to note that this 'doesn't happen magically, but takes time and a structured approach.'

A new data team might be able to run circles around your company's AWS, they have no deeper understanding of what the company's long-term objectives are or even any of the day-to-day challenges that exist directly outside their purview. Likewise, for higher-ups to fully understand the implications of their machine learning decisions, they would likely have to do some kind of online course, which might not be very helpful.

So how do you ensure that your system is providing a structured approach for machine learning to start giving back valuable ROI as quickly as possible?

To answer this very question, Viacom set up a six-month study to look into why certain insights didn't pan out as expected while other initiatives that were successful despite not being implemented immediately. And they came to a very conclusive answer.

Teams which had some form of cross-disciplinary understanding of the other - whether executives who had a greater understanding of the ability and real potential of machine learning or data scientists who had a wider understanding of the company as a whole - tended to do the best.

So they decided to implement an organization-wide initiative to pair data scientists with executives on sandbox projects. These projects, even though the majority of them where never actually implemented, forced the teams to work together with a common goal.

This had to be a well-executed, organizational effort. The teams were led by data scientists with the explicit intention of bringing in stakeholder into the fold.

After a while, two key themes emerged from this 'cross-discipline training'

  • Data teams found themselves having to assess business logic from thought leaders and business insiders. This helped them understand why executive they take certain actions.
  • Business insiders, on the other hand, found themselves in a sort of boot camp. Data scientists helped execs through very basic exposure to math, technology, statistics (which many of them had some basic understanding of already) and decision science.

Marella noted that, 'The response from this initiative was immensely positive'. This is because understanding how the machine space functioned both answered a number of questions and birthed a number of insights.

A particularly interesting discovery were the new capabilities of Excel, which even Sashi admits he was unaware of.

The beauty of this is, while many exec are unfamiliar with machine learning, they are very familiar with how Excel works. This meant they could efficiently come to grips with many machine learning attributes within a familiar program, without having to learn a new computer language.

'By creating well thought out plans and goals where both teams have common of cross-purpose KPI's, with definable dates of completion, encourages the teams to work problems together.' Sashi explained

Having a data scientist on hand to explain what they can do, what they can measure, what kind of models they can implement is invaluable. Stakeholders are capable of implementing those models immediately in their day to day output.

At the end of the day, it really does come down to how you organize your data science teams. This can be:

  • Standalone- This is when the data team is separate from the rest of the decision-making group. The relationship is much more transactional and less visible. Execs approach data teams with a specific task and they prepare it accordingly. However, it is a very ad-hoc process so there can be big communication gaps.
  • Embedded- This, on the other hand, is a much more collaborative affair with deep visibility throughout. Data scientist are responsible for communicating requests to their central teams, meaning less reliance on execs translating requests.

These embedded teams created a highly efficient, multi-perspective units, capable of new heights of innovation. The more data scientists know about overall business processes and objectives, the larger the potential for intrapreneurial prospects.

It also gives rise to more opportunities to generalize processes. Generalization, Sashi explains is 'answering the question of whether a specific code or solution used in one small area of the company, can be scaled up and used to solve more general problems'

So at the end of the day, embedded data teams are any company's best chance at not only unlocking the full potential of machine learning but also employees. Sashi concludes his presentation with the words, ' human decisioning power unfettered by the mundane may ponder more strategic long-term opportunities and visionary pursuits.' 

Bonus Content: Empowering Data Science And Machine Learning: Building An effective Data Strategy 


Read next:

Why Blockchain Hype Must End