The key point for Lucini is this: With so many organisations now rushing to create these kinds of easily accessible services, there is a new imperative for everyone else—to take advantage of the work others are doing to develop the technology, not to compete against them. In other words, rather than try and build core AI applications yourself, let the organisations with the specialist know-how or the cutting-edge research capabilities (such as the Alan Turing Institute) do it for you.
But he also cautions organisations not to be dazzled by some of the more extraordinary things AI is now doing. Take, for example, the DeepMind AI that learnt to play the video game Atari Breakout (in which players use a bat and ball to destroy a wall of bricks). In just a few hours, using the on-screen visuals as its only input and playing millions of games by itself, it had figured out the best strategies and reached superhuman-like levels of skill—a truly amazing feat.
However, this example highlights a key problem in making practical use of AI. How could a business actually use something like this? Lucini’s view is that in many cases, they simply couldn’t. The scenarios just don’t exist. Nobody has 30 million—or even 1 million—examples of anything as suitable for training an algorithm.
Moreover, Dr. Weller explains there are also still technical issues with AIs like DeepMind. That is, they remain quite brittle and have no real idea what they’re actually doing, even if they appear to be doing it very effectively. So, for example, if you change the parameters of the Breakout game just a little (moving the bat closer to the wall, say, or reconfiguring the layout of the bricks) the AI breaks down and has to restart its learning from scratch.
As Dr. Weller set out, this brittleness is just one of the big research challenges still to be tackled—from learning transferable concepts, to accounting for uncertainty, to building common sense understanding—some of which may take decades to resolve. Deep learning, for instance, has proved to be a highly effective form of machine learning. But it’s not perfect by any means. It is, for example, very data hungry and computationally expensive. It struggles to represent uncertainty properly. And it can often be fooled by adversarial examples. So a deep learning algorithm trained, say, to recognise images of animals can be completely fooled if the underlying pixels of an image are shifted ever so slightly towards an image of a different animal. To a human the change will be imperceptible, but the algorithm completely breaks down.
In the end, as Lucini describes it, these challenges mean it’s vital to retain a healthy degree of pragmatism about using AI. But the use cases are there. And in finding them it’s important to try and develop a form of "muscle memory" across the whole range of AI capabilities. Focusing on just one or two (robotics, say, or machine learning) is a risk. That’s because AI’s true value comes from bringing the different pieces of the technology together to develop the use cases that can make a practical difference to today’s organisations.