20.02.2018 Views

NC1801

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

FEATUREAI<br />

AI: A PRACTICAL GUIDE<br />

ARTIFICIAL INTELLIGENCE IS EVERYWHERE, BUT WHAT DOES AI<br />

MEAN EXACTLY AND, CRITICALLY, WHAT OPPORTUNITIES DOES IT<br />

PROMISE? OPENING OUR NEW FEATURE ON AI, PRACTITIONER<br />

PIERS STOBBS, CHIEF DATA OFFICER AT MONEYSUPERMARKET<br />

GROUP, OFFERS SOME PRACTICAL INSIGHT<br />

WHAT IS AI?<br />

You'd be hard pressed to have missed the<br />

acronym AI recently. Artificial Intelligence<br />

crops up in topics as disparate as<br />

healthcare (reading scans), space<br />

exploration (finding planets) and board<br />

games (winning at Go). This very broad<br />

term means different things to different<br />

people, often starting with the spectre of<br />

Terminator-style super intelligent robots<br />

enslaving the human race. But for many<br />

practitioners AI is no more powerful than<br />

the statistical techniques we have used<br />

for decades.<br />

In his thought-provoking book Life 3.0<br />

respected MIT professor Max Tegmark<br />

defines intelligence as "the ability to<br />

accomplish complex goals". He then<br />

defines General Intelligence, often<br />

characterised as human-level general<br />

intelligence, as the ability to accomplish<br />

virtually any goal, with Narrow<br />

Intelligence being restricted to a few<br />

specific goals. Clearly, today's AI<br />

examples are very much in the narrow<br />

intelligence camp. Although there have<br />

been some impressive developments<br />

towards broader capabilities, particularly<br />

by DeepMind and others, the general<br />

consensus is that we are still some way off<br />

artificial general intelligence.<br />

CAN AI HELP?<br />

If today's reality is the application of<br />

narrow AI to specific tasks then there are<br />

many use cases that can build great<br />

advantage. To understand if AI can help a<br />

given problem, characterise the problem<br />

along two dimensions: how obvious is the<br />

answer and how many times do you have<br />

to perform it? If the answer to the problem<br />

is relatively obvious to a human and is a<br />

task that is completed frequently, then it is<br />

likely that AI will help. A classic example<br />

is the amazing progress made by the<br />

Google Brain team in image tagging.<br />

For humans, it is relatively obvious what<br />

objects appear in an image, and we can<br />

easily assign tags. However, at the rate of<br />

a picture every second of every day, it<br />

would take over 30,000 person years to<br />

tag the trillion or so images in Google's<br />

index. Automation might be useful here!<br />

The accuracy achieved by the Google<br />

team using their Deep Learning approach<br />

is astonishing, and it's accessible to<br />

anyone using their API.<br />

IMPLEMENTING AI?<br />

The first thing to think about when<br />

considering AI is data. AI, in this context<br />

supervised machine learning, works by<br />

learning from examples. To build an AI<br />

model you need what is called a data<br />

training-set. Think of this as a really big<br />

spreadsheet where each row contains a<br />

specific example, perhaps information<br />

about a specific customer and the correct<br />

answer such as, did that customer purchase<br />

a specific product? Without a decent set of<br />

data you cannot build an AI model.<br />

You also need to be very clear about the<br />

question you are trying to answer and<br />

how you will act on the model's output.<br />

One popular use case for AI is in churn<br />

prediction to identify which customers are<br />

most likely to stop using your service.<br />

However, it is not always clear what you<br />

should do with the data once obtained.<br />

If you have relevant data and are very<br />

clear about the problem you are<br />

addressing, then it's all about iterating.<br />

Get something simple up and running<br />

and gradually improve it, testing and<br />

measuring all the time. You do not need<br />

the latest and greatest approaches to be<br />

successful. Often it's better to start simply<br />

and establish a baseline to avoid wasting<br />

time spinning your wheels.<br />

Remember to be creative. AI can be<br />

used internally as well as externally. At<br />

Moneysupermarket we have had great<br />

success building an anomaly detection<br />

system to monitor the thousands of data<br />

streams we collect. This is leading to<br />

much speedier detection of potential<br />

issues and is improving data quality. NC<br />

10 NETWORKcomputing JANUARY/FEBRUARY 2018 @NCMagAndAwards<br />

WWW.NETWORKCOMPUTING.CO.UK

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!