EXCLUSIVE: Amazon AI executive explains three things every business needs to address before using machine learning

• Data readiness, business impact and machine learning (ML) strategy are three parameters that every business needs to address before using ML for their needs.
• Data readiness not only includes gathering the right amount of data, but also making sure the non-important bits are filtered out and that customer information is protected.
• Businesses also need to find the right business problem to address — many tend to either undershoot or overshoot.
  

Every business is looking to grab on to the machine learning (ML) bandwagon to either further their own business or narrow the gap against their competition. But before jumping in head first, Vikram Anbazhagan — the Director of Product for Language AI at AWS — believes that there are three critical questions that businesses should ask themselves.

The first of which is data readiness. This includes having sufficient data that’s been cleaner up for ML use. The second is the anticipated business impact — what problem needs to be solved to either boost business efficiency or improve customer experience.


“Many customers know they want to use ML but they’re asking how do I pick the first problem that I need to focus on in order to have a successful ML strategy,” Anbazhagan said at the BI Global Trends Festival 2020.

But, the actual machine learning strategy comes in at the end. The third and final piece of the puzzle addressing how data that a business has curated will match with the impact it has chosen to address.


At the end of the day, ML is nothing more than gathering a lot of historical data, learning from it and then using those lessons to predict what may happen in the future. And, that’s why getting the data part of it right is very critical since it forms the foundation of every use case.

Data readiness
Data is the backbone of all ML applications. Without data, an ML model has nothing to do. “Our customers tell us that more than 50% of their time while building an ML model is spent on getting the data, clearing it and getting it ready for the ML model,” said Anbazhagan.

If the basics aren’t in place, the talent hired to build the ML model will spend the majority of their time just filtering through data rather than building use cases.


In addition to looking at the data that companies already have on hand, they should also plan for the future — start collecting data that may be required a year from now.

‘Data hugging’ isn’t good in the long run
The intrinsic value of data is no secret. Whether it is governments fighting over the data of its citizens or teams within a company.

“We often notice what we call ‘data hugging.’ There are multiple teams in every company working on different projects, and they get really attached to their data — and they don’t want to share their data with other groups,” explained Anbazhagan.

This kind of ‘data hugging’ can work in the short run allowing individual teams to roll out their little ML models, but in the long run, companies need to have an expansive view of all the data across their organisation.

However, while sharing is caring, it’s also essential to have good access control policies and data governance to ensure that any customer data doesn’t fall into the wrong hands. This includes segregating personal data from sensitive data from public data. Any personal details about customers — like address, credit card numbers, contact details, among others — need to be redacted.

Picking a business problem to solve and using ML
This is the tricky part. According to Anbazhagan, some companies tend to either overshoot or undershoot. Ideally, the chosen business problem should have a huge impact, but it should also be something that the ML model can solve within six to 10 months.

“Some companies over-index on the business impact and pick some really complicated problem for which there may not be a lot of data,” he explained.

In other cases, companies may have a lot of data, and the ML model also works well — but the chosen problem isn’t big enough. In such cases, Anbazhagan recommends using the models as proof-of-concept (POCs). This will allow employees to build up their skills and focus on more critical and high impact problems.

Data readiness, business impact, and ML applicability are the three parameters that need to line up for a business to have a successful and robust ML strategy at the helm.

Addressing the ML skill gap
Despite the growing demand for engineers and data scientists that specialise in ML, there is still a shortage in the market.

According to the World Economic Forum (WEF), AI has the potential to create 58 million new jobs by 2022. However, according to Tencent Research, there are only around 300,000 AI engineers around the world.

This makes recruiting talent for ML applications difficult and expensive. Anbazhagan believes companies would be better served by developing internal talent instead and creating a community around reskilling as ML continues to grow.


ML is an inevitable part of the future. It automates most manual tasks that are time-consuming and repetitive in nature — doing them faster and more efficiently. However, it's not something that can be put in place overnight. Putting some thought and direction into a business’ ML strategy can make all the difference between being a leader or a laggard in the industry.

SEE ALSO:
Nokia is set to build the first 4G network on the Moon for NASA — it will eventually be upgraded to 5G

The Moon photobombs NASA's view of the Sun

Wipro wins five-year contract with Europe’s third-largest producer of clean electricity