Intelligent Machines: What does Facebook want with AI?

Google, Facebook and others are busy opening AI labs and poaching some of the most talented university professors to head them up.

Prof Yann LeCun is a hugely influential force in the field of Deep Learning and is now director of AI research at Facebook.

He spoke to the BBC about what the social network is doing with the technology and why he thinks Elon Musk and Stephen Hawking are wrong in their predictions about AI destroying humanity and here are his thoughts.

What is artificial intelligence?

It is the ability of a machine to do things that we deem intelligent behaviour for people or animals. Increasingly it has become the ability for machines to learn by themselves and improve their own performance.

We hear a lot about machines learning but are they really thinking?

The machines that we have at the moment are very primitive in a way. Some of them, to some extent, emulate the basic principles of how the brain works - they are not at all a carbon copy of brain circuits but they have a little bit of the same flavour.

They are very small by biological standards. The biggest neural networks that we simulate have in the order of a few million simulated neurons and a few billion synapses - which are the connections between neurons - and that would put them on par with very small animals, so nothing like what we would think as humans.

In that sense they are not thinking and we are still very far from building machines that can reason, plan, remember properly, have common sense and know how the world works.

But what they can do is recognise objects and images with what seems to be superhuman performance at times and they can do a decent job at translating text from one language to another or recognising speech. So in that sense they do things that humans would consider an intelligent task.

A lot of the machines we are building are artificial neural networks. They do a very large number of simple operations which essentially come down to multiplication and addition.

Those large networks of simulated neurons each of which is connected with several thousand other neurons - some of which come from pixels of an image or an audio signal. It performs a simple operation, something like a weighted sum, of all those values. There are billions of operations to perform.

We organise the neurons in layers - the architecture of this is inspired by visual cortex in mammals - and we can train the machines to recognise objects by showing them thousands of examples.

If we want the machines to recognise aeroplanes, cars, people and tables, we collect lots of images of these things and show them to the machine one after another.

If it gets it wrong, we figure out a way to adjust the strength of the connections between neurons so the next time around it knows what it's looking at.

(BBC)