By Matt Jones
Most articles about artificial intelligence start with big claims. ‘AI will change the world’; ‘Here are some exciting examples of AI’.
These are all legitimate starting points, but fewer words are spent understanding what contemporary AI really is, and who can benefit from it. An uninformed observer could be forgiven for thinking AI is a new technology that you can buy as part of a platform and simply plug into your business and become the next digital company of the future.
Most technology is quite complicated. And AI is ‘quite complicated’ multiplied. If a business wants to really take advantage of AI, they need to stop worrying about what the latest, cool AI gadget or platform will be and start thinking about what an AI toolset can do for the specific problems facing their enterprise.
So, what do we mean by AI?
Firstly, we need to understand what AI is and what it is not. AI is a composite of several techniques and toolkits including; machine learning, deep learning, neural networks, and natural language generation and processing. These tools ingest carefully selected training data to make sense of the task, the information sought and the world in which they live and operate.
What we are not talking about is ‘strong AI’ or Artificial General Intelligence (AGI) – which can, in theory, learn without training, but is still a considerable time from being a reality. This is fine for future gazing, but if you’re looking at what AI can do for you now, put AGI on the backburner.
Building and training AI
If you really want to benefit from AI, you need to develop an AI suited to your problem, and then train it correctly.
One option is to buy an AI black box which will suck in your all your data and spot potentially useful patterns within it. However, many problems worthy of the black box price tag are too complex to automate, and the correlations that pop out the other end are not magically wrapped up in business insight or context, they usually require a lot more work to understand what they really mean, if they’re relevant at all.
Another option is to build the AI yourself. This allows you to bake in an understanding of data and context, rather than using someone else’s approximation. This ensures you understand what’s happening, which helps you identify possible flaws or biases once it’s up and running. And, you get to keep control of all of your data.
There are times when a black box is simpler and we are not advocating building your own AI in every situation, but it certainly offers more control and greater precision of the answers generated. And the good news for people wanting to take this approach is that the leading tech companies—Google, Microsoft and Facebook, etc—have made their AI tools freely available to anyone. These, individually or combined, can be used to build bespoke transformational AI or machine learning platforms, which are as sophisticated as any black box currently on the market, for the price of a data scientist’s salary or consultancy fee.
Customer insights vs operational intelligence
A key part of building and training your AI is understanding what you want it to do.
Most discussion of AI in the media refers to consumer products which aim to guess your behaviour. Much of this is a progression of the data analytics approach used by marketing-led companies like Amazon: mine huge data sets to spot how consumers reacted to promotions, pricing, etc, and use that to predict future behaviour. Finding that a promotion adds 1% to sales is great, it’s not important why the 1% bought or who they were. AI takes this further by using more sophisticated data sets, bringing in new data sources (images, voice etc) and an ability to learn as it goes. But essentially it is still looking for correlations.
Operational and business challenges, on the other hand, tend to use large volumes of data to predict a small number of high-risk events, e.g. under what circumstances a jet engine will fail, a drug will exhibit adverse effects, or oil drilling platform suffer from subsurface corrosion that affects the integrity and safety of operations. They cannot afford failed experiments. They need to know whether one event will lead to another so they can act on which serious money, and sometimes lives, depend.
Doing this needs someone to identify the data needed to train the AI, precisely manage the training, evaluate the outputs and design scientific experiments which can isolate the issue being studied and understand whether one action is the direct result of another – for example, does a change in readings from a jet engine mean it needs a quick clean or that you should take the plane out of service. If you’re not 100% sure, you need to play it safe, which can be very costly.
This needs data skills, the industry expertise to understand what the data means, and an understanding of the scientific method to test patterns, eliminate biases and prove the link between cause and effect. For all the talk of AI as the latest technology, it is as much a scientific issue as a technology one.
Using AI effectively
AI has become a very broad term. The curse of the hype around it is that lots of people are trying to position themselves as experts, shouting about its business changing potential but rarely explaining how.
Platforms with baked in AI has something to offer. Many innovative companies will be developing AI-based products designed for non-expert use. Many of these technologies will be excellent and solve important problems.
But buying an off the shelf product doesn’t immediately solve all your business problems or put you ahead of the competition. If you see AI as an opportunity for your company to make disruptive leaps forward, you need to look at how it can work for you.
AI and machine learning is a series of tools and toolkits. Like any disruptive force, you can’t just buy new tools and expect them to solve your problems. You need people that understand these tools; know how to wield them and when and where they’re most appropriate. That is how to become one of the next disruptors.
Matt Jones is Lead Analytics Strategist at Tessella, Altran’s World Class Center for Analytics.
Original source: Tech Talk