Artificial intelligence: an introduction

Artificial intelligence is all around you and, whether you notice it or not, AI is changing the world. But what is artificial intelligence? And how is artificial intelligence different to machine learning? Can artificial intelligence think like a human? If you are looking for an introduction to artificial intelligence, read on. In this post, I cover what intelligence is, both biological intelligence and artificial intelligence, what kinds of artificial intelligence currently exist and how humans are harnessing the power of artificial intelligence to solve diverse problems.

Related post: how machine learning can help you achieve your goals.

What is intelligence?

There is no single definition of intelligence. Broadly speaking, though, if some entity can extract information from its environment and use it to solve a problem or complete a task, it shows some degree of intelligence. The harder or more complex the problem or task, the more intelligent the entity has to be to succeed, and the more it may require creativity or the capacity to remember, generalise and learn from experience.

Biological intelligence

In the living world, entities as simple as bacteria can interpret information from the environment using biological signal processing. Bacteria have sensors for temperature, light, food concentration, oxygen, amongst others, that feed into a molecular circuit board that outputs a genetically programmed behaviour, like swimming towards a light source. More complex life forms can process more information and exhibit more complex behaviours – making them more intelligent.

• Social insects, such as ants, are well known for organising themselves into successful cities of productivity
• Octopuses have a distributed intelligence: each tentacle has independent decision-making capabilities
• Salmon can spend several years out at sea before returning to the exact same stream they hatched in to spawn
• Whales communicate with one another through song over vast underwater distances
• Humans can decipher the laws of nature and build rockets that send people to the moon and robots to Mars

Humans have also taken the evolution of intelligence and complexity to another level. We now outsource our intelligence to non-living entities.

That is, we create machines that extract information from the environment and use it to solve our problems or perform our tasks. We have endowed non-living entities with ‘artificial’ intelligence.

Artificial intelligence

Like biological intelligence, artificial intelligence ranges from simple to complex.

Basic AI

Let’s start with a simple example: a pocket calculator. This mechanical device receives numbers and arithmetical operators from the environment (input by a human into a keypad) and computes the correct answer.

This may sound like a low level of intelligence, but it can compute far faster than the human brain and solve numerical problems than most people are incapable of.

But it does not require any learning from previous experience: a calculator does not get better at computing if it’s given more practice, and the problems are limited in complexity.

Rule-based AI

Now let’s look at a computer. This is far more advanced. Given a set of instructions (a program), a computer can solve all sorts of complex problems. Using computer programs we can:

– Simulate the Earth’s climate
– Assemble the human genome
– Direct robots in the manufacture of cars
– Visualise the spread of a pandemic
– Use satellites to communicate across the globe
– Control the electrical grid
– Send data from interplanetary space missions

This is all very clever and far beyond human capacity for problem-solving. Without it, our world would be a very different place. The technology is now so common in our daily lives that most of us don’t really consider it a form of intelligence.

But it is.

It’s a form of rule-based artificial intelligence. And yet, it still does not require any form of learning on the part of the machine.

This kind of computer program has to be written by humans. We tell it what behaviour it should perform under what circumstances. We can add some stochasticity to the program but this, too, is hard-coded. All of this makes a computer program, by nature, fixed and completely dependent on the intelligence of the human doing the programming.

Computer programs work exceptionally well for many tasks. But, like the calculator, there are limits to what problems we can solve and how well we can solve them. Because there are limits to human intelligence and the amount of information we can reliably hard-code.

To reach the next level of intelligence, machines have to be able to learn for themselves.

Machine learning

When a computer program learns how to solve a problem, rather than being told how to solve a problem by the human programmer, it’s called machine learning.

This is where the cutting edge of artificial intelligence lies and where the limits for the problems we can solve are expanding rapidly outwards. 

So how does a machine learn?

It has to be trained.

With data.

The method used to train a machine depends on the task being solved and the data available. In each case, the result is simply a series of mathematical equations with weights that accurately map inputs to outputs. Most of the cutting edge AI today uses some form of deep learning, which is a term used to describe using deep (multi-layer) artificial neural networks (algorithms inspired by biological neural networks).

Supervised machine learning

Supervised machine learning teaches by example.

Sometimes humans are able to perform a particular task or solve a particular problem and have created a nice data set on a set of examples in the process. But it is time-consuming, labour-intensive or expensive and we would rather have a machine take over and save us our time, energy and money. Sometimes it’s not even possible for humans to go through the sheer volume of data being produced.

This kind of problem is where machine learning really excels.

You show the machine examples of the correct answers and it learns how to solve the problem itself. This type of machine learning is called supervised machine learning. Once trained, you can then leave the machine to perform the same task for you while you focus on other things. It still requires humans to put in the initial effort and expertise first so that it has something to learn from. Humans also need to keep tabs on the output of the machine to be sure it’s still getting it right – especially if the data are changing over time.

Expert human involvement is critical and yet it is often the most underestimated part of machine learning projects.

So what tasks are we talking about here? What kind of problems can we solve with supervised machine learning?

Solving classification problems with supervised machine learning

Classification is one example of a problem that supervised machine learning works extremely well for if there is a good data set to train on. The process is very similar to teaching humans how to classify things.

Imagine you are teaching a child the names of different animals. As an expert in animal classification, or with the help of a picture book or Google, you show the child photos and tell her the names of the animals. Her brain will learn how to recognise these animals for herself in the future and, after a few months, you no longer need to point and say ‘dog’ every time you see one in the street.

You may want to refine her classification and teach her different breeds of dog using the same process. Then you’ve stepped up her learning. You can do the same thing with a machine and it will learn to solve the same task within hours. And, once trained, a machine is often better at these kinds of tasks than humans.

Biological brains are limited by how fast they can learn. They need time to strengthen connections between neurons, or make new connections.

Machines, with their artificial brains, can adapt in seconds and they can process much more data than a human could in its lifetime.

Supervised machine learning is revolutionising the way humans process information. No longer do we need to feel overwhelmed by the ever-increasing flood of data that our brains, calculators and even computer programs cannot handle. We can set supervised machine learning on the task.

It works for almost any data where you can process a subset yourself and add labels that the machine can learn from. The classification task can be as simple as separating pictures of dogs and cats, classifying the sentiment of movie reviews, or as complex as predicting whether or not an engine part is going to break within the next two weeks. It works on all kinds of data: images, video, text, audio, continuous values of any kind.

Solving regression problems with supervised machine learning

Classification is only one example of how humans can use supervised machine learning. Supervised machine learning can also be used to make predictions with real values. This is called regression. It can make predictions about the future such as the price of houses, the demand for electricity or the remaining useful life of an engine part, by learning from historical data.

Generating new solutions with machine learning

Machines can also learn how to generate entirely new solutions for you. Ones that you may not have thought of and would never have the time to create yourself.

Maybe you want to generate images of people, or convert sketches into photos. Perhaps you have some data but it is poor quality and you want to improve it, or there are some engine fault patterns that are rare in real data and you need more examples for another machine learning task. You might also want to generate text, like captions for photos, or summarise a document.

By showing a machine real examples of what you are looking for, you can train it to produce new examples that look realistic but don’t actually exist.

This is called generative machine learning and it takes artificial intelligence to another level where machines learn elements of creativity.

For now, what machines learn to create will always be similar to the examples they learned from when they were trained. Machine learning has not yet reached the level of creativity where machines can learn entirely new solutions to problems, but this is only a matter of time.

Unsupervised machine learning

With supervised machine learning, you train an algorithm to map inputs to outputs: by giving it examples of the correct answers it figures out how to reach them. You can then apply this trained algorithm to new data and it will give you the correct answers without you spending time figuring them out yourself.

If you don’t have the answers and you think there are patterns or structure in your data that will be valuable or informative if only you could extract them then you can use unsupervised learning. In this case, the machine looks for groupings or clusters that reveal hidden relationships or patterns in your data.

Semi-supervised machine learning

This is a combination of supervised and unsupervised machine learning. It’s very useful if you have a lot of data but only a few curated examples to train a machine using supervised machine learning. First, you train your algorithm using supervised learning on your small data set of examples. Then, you give your trained algorithm the rest of your (unlabelled) data and use its output as input to retrain the algorithm.

Reinforcement learning

Finding the solution to some problems requires making a series of decisions that depend on what’s going on in the world around you. Much of our decision-making is automatic – we don’t think about how to walk around a table, our biological intelligence solves this problem for us in the background by stimulating the appropriate muscles to take the right actions at the right times. We can train an artificial intelligence to achieve the same thing with reinforcement learning.

Think about how babies learn to walk. It’s mostly by trial and error, and reward. For the correct action (taking a step without falling) the child’s brain gets a reward (e.g. getting closer to a favourite toy or enjoying the cheers and smiles of their parents). Each reward reinforces the action they took. They will remember how to repeat that action to achieve the same reward. When they learn to walk they maximise the reward (getting to play with the toy). They will also learn to respond to changes in their environment (like the toy being placed on the other side of the room or a pet cat blocking their path).

Training an AI to walk (either in the real world – in a robot – or in a computer simulation) follows a similar process. The AI senses the environment, tries an action (e.g. making a small movement in one “leg”) and gets a reward back from the environment that tells it how good that action was. The AI tries many different sequences of actions until it learns how to respond effectively in the environment it’s in to achieve its goal.

Artificial General Intelligence (AGI)

Currently, all artificial intelligence is restricted to well-defined tasks. It can achieve better than human performance – but only in the narrow area it has trained on. This is great if you have a specific goal, a good data set and a pre-defined measure of success. You can save yourself a lot of time, energy and money by using machine learning.

But artificial intelligence cannot acquire new skills or adapt to new environments in the way that animal intelligence can: it does not generalise well. How to develop artificial general intelligence is far from clear. It will likely require a new form of machine learning and new ways of measuring the intelligence of machines.

In the meantime, outsourcing our intelligence to machines that can be trained to solve specific problems will continue to shape the future.