Artificial intelligence (AI) and machine learning have emerged as a major talking point for the likes of IBM, Google and Microsoft, with all keen to talk up the work they are doing to help enterprises make the most of these technologies.
Amazon Web Services’s (AWS) has, the company’s CEO Andy Jassy admits, been less talkative about its activity in this space, but that should not be interpreted as a sign that it is late to the AI and machine learning party.
“When we talk to customers and share how we’re thinking about this space, we’ve found a lot of companies don’t realise the heritage Amazon has in machine learning, and part of that is because we don’t talk about it very much. That’s our style,” Jassy told attendees during his keynote presentation at the 2016 AWS Re:Invent customer conference.
As proof of this, Jassy said its parent company, Amazon.com, has been embedding machine learning into its retail site and using it within its fulfilment centres for a very long time.
“One of the earliest features on the internet that people started using was ‘Customers who bought this item might like these items,’ and that was machine learning driven,” said Jassy.
“If you look at how we ask people in our fulfilment centres to pick items, those pick paths are all driven by deep learning and machine learning models.”
Further proof of its work in this area can be found in Amazon’s consumer-focused, voice-controlled Echo device, which interprets users’ vocal commands to play music, retrieve weather reports, and control the lights and thermostats in their homes, he continued.
The device draws on the capabilities of Amazon’s Alexa voice service, which relies on the company’s deep learning, natural language and speech recognition tools to do its job.
“We do a lot of AI as a company and customers say we should talk about that far more,” he said.
Particularly as enterprises are increasingly looking for help making sense of the mountains of big data they accrue during their day-to-day activities.
Artificial intelligence on AWS
This specific area of AWS has been actively assisting customers with through the release of its Deep Learning AMI. This supports machine learning applications running on the Amazon Elastic Compute Cloud (EC2) that can be trained on half a dozen deep learning frameworks.
Chief amongst these is the open source deep learning library, MxNet. AWS confirmed at Re:Invent that this will form the basis of all its AI services in future, having committed just a few weeks ago to helping safeguard the network’s long-term survival through code contributions and supporting tools.
The company’s high-performance computing-focused EC2 P2 instance family has also emerged as a popular location for customers to carry out their machine learning work, added Jassy.
“We’ve released a whole number of pieces that have led to a lot of customers doing this type of work in AWS. The P2 instance family is really optimised for deep learning and we have an unbelievable number of customers doing deep learning on top of those instances. It’s one of the fastest growing instances in our history,” he said.
In terms of reference customers, Jassy cited Fraud.Net. Its fraud detection and prevention platform draws on Amazon’s machine learning capabilities to analyse payment data in real time to pick up on suspicious activity.
For the merchants that rely on Fraud.net’s services to prevent them from falling victim to fraudulent payments, Jassy said the service is collectively saving them around $1m a week.
Another example is not-for-profit Thorn, who used Re:Invent to share details of how it is using machine learning and AI technology to support law enforcers in their work to track down sex traffickers across America, including those who use internet sites to advertise underage victims.
The company’s Spotlight app processes 200,000 adverts a day from these sites, to help law enforcers establish links based on the data they contain in order to identify traffickers, and locate and protect their victims.
Before Spotlight, law enforcers would be forced to manually scroll through the adverts, recording data they think might be significant. But with adverts often disappearing within hours of being posted, and with the perpetrators going to great lengths to obfuscate phone numbers and other key data, it could be difficult for them to keep up with the perpetrators.
“What law enforcement would do is they go onto the internet and start writing things from these sites that they think might be related. This is not how you can tackle the problem at scale,” said Kristin Boorse, senior product manager at Thorn, during a customer session at Re:Invent.
“We know, based on our surveys, that for law enforcement who are coming in on a daily basis, they’re saving 60% of their investigation time because we can show the relationships between the data that is there. It’s meant significant time-saving for law enforcement.”
Speaking to Computer Weekly at Re:Invent, Gavin Jackson, managing director of AWS in the UK and Ireland, cited travel search application Skyscanner as another long-time user of its machine learning tools, which it has embedded in its services so users can book flights using voice commands.
Connected car manufacturers are also emerging as keen adopters of the technology, he added.
“If you think about the experience of driving a car now and the distractions you have from technologies, such as mobile phones, if you have voice activation in all these things it helps a lot,” he said.
“It’s no surprise to say that sort of intelligence using AWS technology is very popular with next-generation connected cars. BMW have launched it and there are a number of others coming down the pipeline. We’re seeing these technologies being pervasive in everything.”
Accessible to all
The machine learning and AI services AWS has provided customers with to-date have been primarily aimed at “expert practitioners”, said Jassy.
However, the company is now working on putting this technology in the hands of everyday developers through the rollout of three new services under the Amazon AI brand, with more to follow next year.
All three, unveiled at Re: Invent, are designed to enable developers to embed aspects of machine learning and AI into their web, mobile and connected device applications.
They include its text-to-speech conversion service Amazon Polly, an image analysis tool Amazon Rekognition and Amazon Lex.
The latter features the same technology as the aforementioned Alex, and is geared towards developers that want to build conversational user experiences into their applications.
Jackson said the company’s decision to bring these products under the Amazon AI brand should help ensure they get on the radar of customers who may have previously been oblivious to the work the company is doing in this space.
Read more about machine learning and artificial intelligence
- As the idea of using artificial intelligence to reduce energy use by datacentres gathers momentum, does this spell the end of human intervention on sustainability matters?
- Ocado has revealed details of how the deployment of machine learning technology is boosting the responsiveness of its customer services team by helping them avoid email overload.
“The fact we’re extracting all of our past experience and intelligence and putting it into a product set, adding other things to it from other developments we have, such as Alexa, really brings it to life. It should help our customer digest all of the messaging we have on this,” he said.
Speaking to Computer Weekly at Re:Invent, AWS chief evangelist Jeff Barr said Rekognition has been trained on “billions of images” to ensure any photos it is tested on are identified with a high degree of accuracy.
Where the service looks set to come into its own is where users have huge libraries of images to manually sift through where the people or items contained in them need to be swiftly identified.
Barr said he knows of customers who are using the technology to protect access to sensitive corporate data on devices by concealing it should someone walk into view who is not authorised to see it.
There are lots of uses cases for the technology, but building these capabilities into applications can be labour and resource intensive, making it difficult for companies without the required technical expertise in-house to reap the benefits.
“This is problem a lot of customers face. They’re presented, with everyone having cell phones and uploading pictures to places, requiring automatic categorisation, automatic identification, but if you have a list of faces, you can train recognition on that, but it’s a problem that’s very difficult to solve on your own,” said Barr.
“It’s technically challenging, requires a lot of storage, a lot of specialist knowledge of deep-learning, image processing and feature extraction, and all of that is packaged up and delivered to customers in the service.
“So instead of PhD-level knowledge and building something custom, they can add image analysis and recognition to their existing applications in days,” he added.
AWS’s Jackson said the company’s commitment to lowering the barriers to AI and machine learning for developers could transform the way many applications are built in the future.
“Writing artificial intelligence and machine learning natively into applications is going to be the new normal. By default, you’re going to have these capabilities written into applications in the way that social has been written into applications as well,” he said.