Former US intelligence officer Drew Perez is an old hand at making sense of vast volumes of data using machine learning and artificial intelligence (AI) in the name of counter-terrorism and national security.
Armed with this know-how gleaned over 30 years, Perez founded Adatos, a Singapore-based AI startup where he adapted recently declassified methodologies and software used by US defence and intelligence communities to produce data-driven insights quickly.
“We’ve been doing this in the intelligence community for decades,” says Perez. “There’s really nothing sexy about it, because it just works.”
The intelligence community first discovered machine learning during World War II at the UK Government Communications Headquarters at Bletchley Park, where the famed German Enigma codes were broken using techniques that laid the foundation for computing and AI.
Since then, Perez says intelligence services and the US military have been using machine learning and AI to process vast amounts of data, with previously unmanageable signal to noise ratios.
“The current demands in counter-terrorism require precise, accurate insights delivered in the span of minutes,” he says. “A key contest in war will be between adversary cognitive systems – artificial and human – to process information, understand the battlespace and decide and execute faster than the enemy.”
While AI and machine learning are not new to military and intelligence communities, the world has been enamoured by their ability to beat chess masters and human opponents in TV game shows over the past decade.
Much of these developments have been fuelled by rapid advancements in computing power – think Moore’s Law – and large volumes of data being generated by sensors and mobile devices.
However, Perez says the hype and expectations around AI and machine learning today may lead to disappointment, if people are expecting Ex Machina-type humanoids with the ability to think like humans do.
“AI, if defined by the expectations of cognitive functions that mimic humans, is still largely in a development stage, but it doesn’t mean it can’t solve real-world problems much more efficiently,” Perez says.
Practical applications in APAC
Machine learning as a form of AI is already widely accessible for immediate practical applications.
It has rapidly matured over the years, and is the norm in a number of fields, helping companies deploy smart systems of engagement to improve efficiency, enhance security, gain insights and deliver superior customer experiences, says Aan Chauhan, chief technology officer at Cognizant.
According to research firm IDC, the Asia-Pacific AI market will grow at a compound annual growth rate of 63.9% between 2015 and 2020, surpassing the global growth rate of 55.1%.
“Be it virtual cognitive agents delivering contextual and personalised services and customer interactions, predictive analytics engines that help companies forecast, systems that help automate business processes, applications that automate infrastructure management and application services, or deep learning systems that augment human expert capabilities, machine learning applications are being used across industries,” says Chauhan.
Examples include robo-advisors in financial services, sales forecasting in retail, supply chain optimisation in logistics, robotic process automation and even medical image analysis, such as screening images of the retina for diabetic retinopathy.
“AI and machine learning platforms take a while to ‘learn’, but the effectiveness of the engine improves with time,” says Chauhan.
Tracy Tsai, research vice-president with Gartner’s personal technologies team, says for now, repetitive tasks – as well as jobs that require humans to process large amounts of information to make decisions – are most suited for machine learning.
“For example, it is difficult in crowded areas for the human eye to recognise faces of criminals amid so many moving objects,” she says. “That’s where machine learning and AI can be used in surveillance applications.”
High-fidelity machine learning
In cyber security, machine learning enables security software to recognise known threats that are attacking a system, and subsequently stop them from passing through.
This process also helps the software build up a threat database, allowing it to recognise and block more threats as time goes on, says Ryan Flores, senior manager for forward-looking threat research at Trend Micro, Asia-Pacific.
Traditional machine learning, however, becomes inadequate when it comes to unknown threats that a system has never encountered before.
“Traditional machine learning-powered security software is only able to run tests before a file’s execution and, in some cases, allow the threats to pass through. This approach is outdated as many threats only show their malicious intent on execution; it’s too late if it has entered the system by then,” says Flores.
A better approach would be what Flores calls “high-fidelity machine learning”, which runs tests before and during malware execution, allowing security applications to halt malicious operations halfway and kick malware out of the system.
High-fidelity machine learning also enables many other defence capabilities, such as reducing false positives, behavioural analysis and preventing operating systems from being exploited.
Flores says such machine learning capabilities can autonomously and intelligently choose the right defence for the corresponding threat, easing the computational load on IT systems.
Descriptive machine learning aids marketers
Marketers are also harnessing machine learning to better predict how certain customers react to various marketing efforts and how likely they are to make a purchase in what’s known as conversion.
This helps brands and agencies run more holistic marketing campaigns, targeting the right audience through the most optimal channels and at the best times, says Heather Blank, senior vice-president at MediaMath.
However, Blank says machine learning is not necessarily always predictive in nature and can be descriptive instead.
“Studying machine learning models can help explain which features or individual characteristics are important in predicting an event – usually a conversion – and which features may be meaningless or even predictive of the event not occurring at all,” she says.
“This can help marketers understand their consumer patterns more clearly, by filtering out the noise. It can also help challenge status quo notions of what is important to the purchase cycle or even what an ideal consumer looks like.”
Over-reliance could have significant consequences, says expert
Cognizant’s Chauhan notes that rapid adoption of machine learning has also brought some of its limitations to the fore. These include the presence of data in silos, limited availability of deep data analytics skills, varying accuracies of algorithms and the speed at which things are changing.
As technology matures, Chauhan warns that an over-reliance on machine learning or a misunderstanding of its abilities could have significant consequences, especially since machine learning-based applications might not be able to fully comprehend peculiarities of human sentiments and cultural contexts right at the outset.
“Businesses must realise that machine learning is primarily designed to help employees get better at what they do, and not as a tool to replace people,” he says.
In some cases, however, machine learning and AI can replace data scientists, says Adatos’ Perez.
“Rather than using data scientists to build risk-based models in the financial industry, we’re using machines to do so at a much faster rate,” Perez says, adding that regulators would have to come to terms with such models.
“Because machines are building these models so fast and accurately, regulatory bodies would either have to get enough experience to test and trust the machines, or act as a constraining force on something they may not fully understand,” he says.
Getting started on machine learning and AI
As machine learning and artificial intelligence technologies become more accessible, organisations will need to be prepared for their impact on the workplace, says Gartner’s Tsai. “We will see a lot more cases in the next 10 years, so companies will need to get started to remain competitive,” she says.
Trend Micro’s Flores says organisations will need to identify key areas that require machine learning. For example, do they need more customer insights? Do they want to know more about how their product is being used? Is there a way to improve the product or service they offer? Asking questions that matter to their business operations will assist companies in knowing what they need machine learning for.
But not all companies will have the talent and skills required to build data models and integrate the technologies required for AI and machine learning. “They should turn to suppliers with proven track records in specific industries or specialised areas such as natural language processing,” says Tsai.
Ultimately, machine learning is a field where the devil is in the details, where it is very easy to create something that seems correct, but is not, says MediaMath’s Blank.
“They should also understand that machine learning is a powerful tool, but it has its limitations – it is not a magic box. It can solve many problems, but it can’t solve them all, and one of the best determining factors of its success is the quality, relevancy and freshness of data being used in the first place,” she says.
Read more about machine learning and AI
- Computer Weekly finds out how Bloomberg’s Gideon Mann balances the certainty of analytics with the alchemy of machine learning.
- Domino’s Pizza launches a chatbot that takes orders via its Australian website and mobile app.
- Fei-Fei Li, Google’s chief scientist of artificial intelligence and machine learning, talks up societal benefits of making artificial intelligence tools and technologies accessible to all.