Brain (2)

Artificial Intelligence: it conjures up an image of thinking machines – but what is it really? And what is deep learning, machine learning and what are neural networks? The Amazon CEO has come up with a neat explanation of machine learning and AI, and we have added a few more.

To put it all in a nutshell, AI refers to machines acting in a way that appears smart. You could say that machine learning refers to where AI is right now, and deep learning is a subset of machine learning. Neural networks describe a way AI can occur.

In times like these, when jargon gets thrown around left right and centre, we need a bit of clarity, and who better than Jeff Bezos, founder and CEO of Amazon – a man who is within a whisker of becoming the richest man in the world, to explain it.

In his latest letter to shareholders, this is how Mr Bezos put it:

“Over the past decades, computers have broadly automated tasks that programmers could describe with clear rules and algorithms. Modern machine learning techniques now allow us to do the same for tasks where describing the precise rules is much harder.”

And then he drills down:

“But much of what we do with machine learning happens beneath the surface. Machine learning drives our algorithms for demand forecasting, product search ranking, product and deals recommendations, merchandising placements, fraud detection, translations, and much more. Though less visible, much of the impact of machine learning will be of this type — quietly but meaningfully improving core operations.”

But let’s pull back.

At its simplest level, AI uses rules – for example if X, then Y happens. Deep Blue, the machine that was developed by IBM and which beat the world champion at chess, uses AI.

DeepMind, the machine that won the Chinese game Go, and was developed by Alphabet/Google subsidiary DeepMind Technologies, applies machine learning techniques.

Machine learning is an example of convergence – it combines the technology for allowing machines to learn, with the data available from the internet. The concept is not new, it goes back to 1959 when Arthur Samuel defined it as giving “computers the ability to learn without being explicitly programmed." However, being able to take data from the internet and learn from it: that is close to being state of the art

Machine learning makes use of neural networks.

The brain works in different ways to computers – the brain is made up of over a hundred billion neurons that link to each other forming synapses. To use the jargon of network theory, the neurons in the brain are called nodes, they form a network, and some nodes have more connections than others, they are called hubs. A neural network mirrors this – at least it does so in a simplified way. A neural network consists of artificial neurons.

The concept of neural networks is not new, but what is new is the technology to make such networks massive in scale, until this technology existed, neural networks were ineffective, but technology passed a kind of tipping point around the beginning of this decade. In 2012, for example, Andrew Ng, at Google, managed to apply a neural network to identify an image of a cat.

Machine learning can learn from data fed to it via neural networks.

Okay, so what’s the difference between machine learning and deep learning?

Deep learning takes the above concepts a stage further – it makes far more sophisticated use of neural networks applying multiple layers of nonlinear processing units, mimicking the way the human mind works to make decisions.

Deep learning is the technology that makes autonomous cars possible.

But we have already hit worrying problems with deep learning. Because of the way it learns, it can reflect human biases – you may have read in the media about racist algorithms for example. As Joanna Bryson, a computer scientist at the University of Bath which conducted research into AI reflecting bias said: “A lot of people are saying this is showing that AI is prejudiced. No. This is showing we’re prejudiced and that AI is learning it.” But she added: “A danger would be if you had an AI system that didn’t have an explicit part that was driven by moral ideas, that would be bad.”