DEEP - LEARNING DIVIDEND

Even small firms and startups are beginning to use advanced deep-learning tools and algorithms, says Amazon Web Services’ machine learning chief.

I believe that in the future, every application is going to have a machine-learning component already built in

With the cloud and machine learning, [businesses] can quickly set up contact centres literally in a matter of hours, and they can build a chat bot to engage with customers and spread information quickly

Swami Sivasubramanian, vice-president of the machine learning division of Amazon.com, sees opportunities everywhere — even amid the fatigue of jet lag.

While on a month-long family vacation in his home country of India four years ago, a thought kept bugging him. The world was on the verge of seeing deep-learning technologies becoming mainstream. But this was not going to happen until someone connected the dots to make it easy for any company to adopt these technologies. At the time, they were mostly accessible only to top companies and research institutions.

So a jet-lagged Mr Sivasubramanian, who was then the general manager of NoSQL (a database) and analytics for Amazon Web Services (AWS), began to spend many of his waking hours teaching himself the different facets of deep learning and finding out where the gaps were.

When he returned to Seattle a month later, he wrote a six-page presentation that outlined the need to accelerate the adoption of machine learning for a broad variety of developers, and how Amazon could play a big role in this. Mr Sivasubramanian presented his idea to Amazon’s leadership — who promptly put him in charge of the new business.

Today, most machine learning in the cloud is done on AWS. It launched more than 250 machine-learning features and capabilities last year alone.

The company’s most prominent product, Amazon SageMaker, helps tens of thousands of active customers to quickly build and train machine-learning models in the cloud. It has been used for for everything from predicting congestive heart failure, analysing NFL games and matching Fender guitar wood with the best shade of paint, to predicting anomalies in space that might hint at a solar storm.

NEW GOLDEN AGE

Ever one to glimpse opportunities, Mr Sivasubramanian believes the world is witnessing the dawn of the golden age of machine learning and artificial intelligence (AI).

“There is so much innovation happening right now, and more innovation is about to happen. In fact, I believe that in the future, every application is going to have a machine-learning component already built in,” he says.

That’s an exciting prospect to ponder, but AI attracted a surge of interest three decades ago before descending into its figurative winter — a cycle that has since recurred over the past few years. Why would anything be different now?

For one, the access to computing infrastructure and data storage has now been democratised, Mr Sivasubramanian says.

“There is a class of machine-learning algorithms called deep learning,” he explains. “Even though the original idea behind deep learning was published 30 years ago, at that time the deep learning algorithms learned from huge amounts of data to identify the patterns to build a speech recognition system, or to build personalisation or image classification and so forth. It required a huge amount of computing resources and huge amounts of data.

“And this turned out to be not viable during those times. If you’re one of the few research labs in the world, or one of the big technology organisations, maybe you get further. But it was not affordable for a two-person startup in a garage.”

The new-found mass access to computing and data storage capabilities has changed everything.

“Suddenly, now you, as a startup developer, will be able to pay by the hour for computing and storage, and actually build this amazing infrastructure,” says Mr Sivasubramanian.

“This was one of the biggest changes — that’s why you now see that these state-of-the-art deep-learning infrastructure and algorithms are starting to get used, not just by big organisations, but also everyday startups.”

The versatility of Amazon’s machine-learning tools is reflected in the diversity of startups that have benefited from the trend. For example, the Malaysian startup Amazing Fables, which creates personalised children’s books, uses Amazon Polly to integrate text-to-speech functions in 10 languages for its video books.

HappyFresh, an Indonesian retailer backed by the Southeast Asian super-app Grab, uses services such as AWS Lambda to support fraud detection and Amazon SageMaker to personalise the online grocery shopping experience.

Large organisations have also used machine learning to improve their internal processes and customer offerings. The California-based financial and tax software giant Intuit has developed models that can pull a year’s worth of bank transactions and find deductible business expenses for customers. It also managed to reduce machine-learning deployment time by 90%, from six months to one week, by using SageMaker.

Meanwhile, data scientists at Formula 1 are training deep-learning models with 65 years of historical motor-racing data to extract critical race performance statistics. During each race, 120 sensors on each car generate 3 gigabytes of data, with more than 1,500 data points created per second.

The real beauty of machine learning today is that one no longer has to be an erudite techie to be able to build machine learning-driven systems, Mr Sivasubramanian says. The wide variety of tools available allows anyone, from insurance companies to e-commerce retailers, to build systems that help serve customers better.

PANDEMIC PUSH

Innovative technologies are playing crucial roles during the Covid-19 pandemic. They have helped people connect with each other amid global lockdowns, kickstarted contact-tracing programmes and much more.

AI and machine learning have also played an active role in helping the world to understand and better address the Covid-19 crisis in both the public and private sectors.

“Now, more than ever, machine learning is going to be such an important technology to help us during this pandemic,” Mr Sivasubramanian emphasises.

In the medical field, for instance, experts are building algorithms for earlier detection of respiratory diseases linked to possible Covid-19 infection. In the United States, UC San Diego Health trained its machine-learning model by inputting 22,000 notations by radiologists to help with quickly diagnosing pneumonia through chest X-rays.

And with the world in a race against time to find treatments for Covid-19, companies such as AstraZeneca and UK-based BenevolentAI are tapping machine learning to accelerate drug discovery.

Companies in various industries are also finding ways to keep their customers close to them amid the pandemic. “With the cloud and machine learning, they can quickly set up contact centres literally in a matter of hours, and they can build a chat bot to engage with customers and spread information quickly,” Mr Sivasubramanian says.

Many businesses are also rebuilding their supply chains using machine learning to do better forecasting, so that they can scale, adjust and adapt to changing conditions, he adds.

“The Covid-19 pandemic has made many organisations realise that digitisation and digital transformation was not a programme you had to postpone.”

AI technology, in particular, is unlike many disruptive technologies that usually have a pattern of gaining traction in one or two industries before going mainstream. AI has quickly burgeoned from being exclusively adopted in cutting-edge industries to its current widespread uptake by almost every industry segment, Mr Sivasubramanian points out.

But more can be done to apply machine learning in businesses, he believes. Industrial and manufacturing applications in particular could benefit from further development.

“There is a lot of opportunity in the space of predictive maintenance that could leverage machine learning, instead of customers trying to do constant maintenance on a regular schedule, or incurring downtime.”

COMMON MISCONCEPTIONS

While machine-learning technologies are on the rise, some misunderstandings still impede widespread and effective adoption. People most commonly fail to understand that the models need to be constantly refined.

“It is not a one-time do it and then you’re done,” Mr Sivasubramanian says, with a wry smile that betrays how often he has had to explain that.

“You don’t build a machine-learning algorithm, and then say, ‘Now that I have built it, it’s done and we’re over,’ because the workload patterns change and customer behaviour changes. And you have to constantly invest and continue to improve the models.”

Unlike standard application development, machine learning requires that one constantly iterate and improve the accuracy of the model. “So I encourage leaders to not be deterred by the initial results — if it is only 40% accurate, then they say machine learning won’t work for it. It is one of those things where you continue to investigate and iterate and improve.”

Leaders also commonly fail to realise that data is the fuel for AI. Many companies invest in top-tier talents who command eye-watering salaries, but not enough in their data strategy.

“And this is where AWS has invested quite a bit,” he says. “We want to show that they can actually build a highly scalable, secure and reliable data lake (a repository of raw data) so that they don’t end up spending half of the time data wrangling instead of writing machine-learning algorithms.”

Given his background and his role, Mr Sivasubramanian’s preferred advice is surprisingly human — focus on the customer first. That is, solve the problem that only your company or organisation is uniquely positioned to solve, he says.

In fact, if it were possible for him to give a piece of advice to his younger self who had just joined Amazon as an intern in 2005 (while completing a PhD), this would have been it.

“I realise that, more than technology, the moment you listen intensely to customer feedback and then you use technology as a means to serve their needs, you will actually accelerate your own personal learning and also meet customer needs and grow the business,” he says.

The other piece of advice he would give his younger self? “Don’t ever sell Amazon stock.”