1. What is Intelligence?

We have a brain that helps us understand the world around us, we learn as we go along although somethings we have inherited, pre-built or pre-programmed. A child learns to crawl and then walk, it learns about the difference between a cat and a dog. We store memories and apply reason, and learn how to solve problems we haven’t encountered before. To do all this we use our brains which are full of billions of Neurons that fire continually sending electrical impulses.

Most Artificially Intelligent (AI) machines have something called an Artificial Neural Network. This was created to imitate the human brain. There aren’t any Artificial Neurons just some clever programming and maths to simulate how the brain might work if it was a machine. We train the AI machine by providing the data, we then tweak the Artificial Neural Network to work better.

A good analogy is the difference between cats and dogs. A young child will very quickly and very precisely work out the difference between a cat and a dog. Although it may not live with a cat or a dog it will start to identify one or the other very accurately very quickly. It will start to know what a cat or a dog is and how they behave, sound etc.

An AI machine is trained on thousands of images of cats and dogs. Over time it can predict with reasonable accuracy the difference between a cat and a dog. You may think that if that is the case then you could train it to do anything. The problem is it still doesn’t actually know what a cat is or a dog is. All it has done is learned to recognise to a degree of certainty whether the image is of a cat or a dog.

As humans we look for patterns. Watching a baby learn to crawl is fascinating it tries all kinds of movements of arms and legs before finally working out the right sequence of movements that will propel it forward to reach the toy that it wants.

Training

Every AI model has to be trained on some data. That training can take minutes, hours, days even weeks depending on the amount of data and the size of the model.

Data Centres

Data Centres are where the model is trained but also where the model is deployed called inference. This is can mean where it predicts a particular outcome depending on the purpose and actions that the model is trained for.

2. Data is everything

An AI machine looks for patterns and then predicts. Some call it an approximation machine. When categorising cats and dogs it is looking at the features it has learned to recognise, the shape of the ears, nose etc, however to do that it needs lots and lots of data, 10’s of thousands of images.

We do this as well, we recognise people’s faces, we may have only met them once but we would’ve sent multiple images from different angles to our brains, also their voice and movement. Our brains are brilliant at processing all of this information, labelling it and storing it for future use.

What seems relatively effortless and easy for a human is actually very hard for an AI machine. Remember that it is just some wires, a program with some fairly straightforward mathematics. You have to show it an image of a cat and tell it that it is a cat, same with a dog and so on. The program is tweaked every time it sees a labelled image and goes onto the next image.

You might do this with a child, asking them what a particular animal is in a picture or story. If they get it wrong you tell them and over time they start to even recognise drawings and painting of animals. Something an AI machine might very well struggle to interpret.

Good data is critical for machines to learn. One example is when an AI machine was learning to recognise different animals. All the images of cats and dogs were taken with a background that was either inside a house or outside in the garden or countryside. When it was tested it thought a dog was a bear. That was because the image of the dog had a snowy background and so a white background meant it must be a bear.

Data needs to be varied, accurate (label everything) and have plenty of it. Life, however is never that neat and that is where unforeseen problems arise.

3. What is inside an AI Machine?

The majority have what is called an Artificial Neural Network, or Neural Network for short. A Neural Network is not a machine in itself. It is an algorithm which is a posh way of saying it is a collection of instructions. Algorithms are used in all coding, with AI it is one particular type of algorithm which isn’t as complicated as you light think.

You could, with a bit of practise, make one from scratch. It may not be able to do much but it would be an AI model. However, most are far more complicated than that because they also have to be efficient as they process vast amounts of data, but the principles are simple enough. The maths is no more than a bit of High School level maths.

The great things today is that all the hard work is now done for you. These algorithms are now packaged so that you just need to make a few adjustments. These models or Neural Networks are extremely powerful in what they can do and anyone can have a go. Yet they aren’t the solution to everything as some may think, they don’t always work very well. Training a model is more of an art than a science.

We can explain what a neural network is through a diagram. This is where the language can be a bit confusing if you are not familiar but it is the best way to describe it. The human brain was the inspiration for the development of AI back in the 1950’s. There are lots of elements to a neuron in the brain but at the basic level there are a number of electrical inputs which can trigger an output which fires electric pulse.

4. The Perceptron

This is a perceptron is very limited on its own. It can be used to solve linear problems, ie straight line solutions. An example is drawing all the points about a line red and those below the line blue. The inputs are the co-ordinates and the output is the colour.