Tap into powers of big tech companies

What if you could tap into the same level of artificial intelligence (AI) that the most advanced consumer internet companies use to run their services?

The answer, apparently, is you can. That, at least, is the promise behind some of the “machine learning in the cloud” services that are now becoming widely available.

Amazon Web Services (AWS), for instance, added machine learning — the ability for computers to learn without programming — to its list of available services last year. This gives customers the ability to apply the same algorithms to their own data that Amazon has used internally for fraud detection and to drive its website recommendations.

The online retailer had set up a centralised machine learning unit to develop the technology and make it available to internal developers, says Matt Wood, general manager of product strategy at AWS. It was then relatively straightforward to open up the service to outside customers.

Companies that store their data in the AWS cloud can now build AI models and train them using data they supply. These models can be used for predictive tasks such as forecasting demand or anticipating how customers will behave in certain conditions.

“Once customers start to apply predictive models to their data, it becomes addictive,” says Mr Wood.

IBM, through its Watson “cognitive computing” division, also provides AI services. The platform uses natural language processing and machine learning to interrogate large amounts of clients’ unstructured data.

AWS’s size means it can offer low unit costs while handling the large amounts of computing power that are needed to run such predictive data models, Mr Wood says.

When it comes to the most advanced forms of AI, however, Amazon is a “junior player”, according to Tom Austin, an analyst at Gartner.

He says Google and Microsoft are further ahead in the field of deep machine learning, as are Facebook and Chinese company Baidu. And it is deep machine learning — a technique that uses so-called neural networks, modelled on the human brain — that could offer the greatest potential.

Jeff Dean, a lead engineer at Google, says neural networks can reach conclusions automatically by trawling through large volumes of raw data without needing the sort of human guidance standard machine learning algorithms require.

Turning this into a practical, everyday business tool poses challenges, however. One is time. The amount of computing required to train the technology on a new data set can be a barrier to practical use. At Google, engineers working on such exercises hit the “patience threshold” where frustration sets in after around four or five days, says Mr Dean.

A further difficulty lies in making the technology available in a way that it can be applied easily by other developers to their own problems. “How we can package that up and make that easier to use is a question we’ve been wrestling with within our group,” Mr Dean says.