Are You Smarter Than a Neural Network?

Last week I was invited by the resourcing department of a large international bank to give a talk on The Future of Artificial Intelligence. We soon entered into an animated discussion on how clever artificial intelligence or AI algorithms really are. Let's focus on artificial neural networks for a moment. Artificial neural networks are by no means the only AI method in use today. However neural networks are perhaps the technique that receives the most contemporary press and is guiding - and simultaneously misinforming -  the debate about AI. 

Artificial neural networks are loosely modelled on the human nervous system where an input (e.g. a telltale smell of smoke) is translated into an output response (e.g. sound the alarm). This happens when the input breaches a certain threshold. 

Crudely put, the more neurons a nervous system has, the smarter it appears to be. As illustrated below, a human brain is much more evolved than the brain of a cat. This allows humans to do things a cat could never do. And with the exception of William Hanna and Joseph Barbera’s Tom and Jerry cartoons, a cat is in turn, smarter than a mouse, and a mouse smarter than a cockroach, which is smarter than a fly. The number of neurons plays a large role in this hierarchy.

Similarly, the more intricate an artificial neural network’s topology, or the more neurons it has, the more complex the problems it is able to tackle. These range from classifying mushrooms as edible or poisonous from a set of input characteristics, to managing the energy grid of an entire city.

When asked how many neurons state of the art artificial neural networks typically have today, the audience’s responses from last week’s talk ranged from a bullish, ’many more than a human’ to a tentative, ‘somewhere around a fruit fly’. It turns out that the latter response is closer to the truth.

For all their enormous versatility and promise, AI algorithms today perform relatively simple tasks. These include image recognition, text translation and even piloting autonomous cars. The AI that classifies mushrooms as edible or poisonous, has no notion of what a mushroom is. It simply takes a set of items with certain attributes and labels them as ‘edible’ and labels another set of items with a different combination of attributes, as ‘poisonous’. A truly sentient machine, one that falls in love, laughs, experiences sorrow and can be driven to great heights of achievement by hope alone, is still a long way off.

So what is the future of AI? This is somewhat analogous to the central question in The Hitchhiker’s Guide to the Galaxy by Douglas Adams. That question of course, was, ‘What is the answer to the ultimate question of life, the universe and everything?’ The answer to that question was, ’42’. Or in other words, nobody really knows.