A practical definition for business leaders: machine learning is the training, AI is the decision making based on that training.
Much of that content is of poor quality, seemingly copied and pasted from the same sources. Some of it is actually misleading.
Despite this, it is worth recognising that there is a genuine need that attracts these promises of clarity.
To make matters worse, the terminology around AI has been used in different ways through history, and this continues today even amongst experts.
A more useful approach is to consider which definitions are going to be most useful to business leaders who are trying explore the opportunities of AI, clarify its operational impact, and understand the risks they need to own and manage.
In that context the following definitions are useful:
Not everything should be labelled machine learning or AI, even if marketing teams are furiously trying to attach the labels to everything they can.
A simple test is whether the AI has achieved a level of competence comparable to a human expert in a given task. If the machine learning process requires a human expert to encode lots of his or her knowledge into the process, then it is not really machine learning, it is simply a translation of that knowledge into computer instructions.
If a person was to do this task it would be slow, laborious, and because people get tired, there is an increasing risk of error too.
We could try to get a computer to do this sorting. A computer doesn't get bored, nor tired, and modern computers are very fast and can crunch through a million photos within minutes.
We could teach a computer to tell the difference between a cat and a dog by writing down detailed rules with the help of an expert, a zoologist in this case.
This might work. We can describe rules that ensure the animals have two eyes, four legs and a tail. To separate cats from dogs, our rules might say that cats have different proportions for head and leg size to dogs. This is more difficult to get right.
In fact, even writing rules to find four legs in a photo is difficult. Some legs might be hidden, or the angle might make it hard to distinguish them.
This approach of trying to write very detailed rules doesn't generally work well. The the rules themselves can be difficult to get right, the process of preparing them is hard, and there are many scenarios where we don't know the rules in sufficient detail.
An alternative approach is to see if we can get a computer to work out the rules which separate cats and dogs by itself. This is machine learning.
The computer learns by being shown lots and lots of labelled cat and dog photos. Each example tweaks the computer's understanding a little bit. Initially the computer's understanding is non-existent, but after many training examples, the computer's understanding of what makes a photo a cat or a dog has improved.
You can see parallels between this gradual training process and pavlovian learning.
Once we've trained the computer, we can start using it to classify photos that it hasn't seen before. This automated decision making, based on a previous machine learning step, is called artificial intelligence, or AI.
In practice, not all attempts at machine learning work.
And even when they do, we need to carefully monitor how well the computer performs on unseen data. It may be that the accuracy can be improved with more training, or it may be that our particular architecture will never reach the desired accuracy.
This is impressive enough.
But the real power of machine learning and AI is demonstrated when this can be done more accurately than a human expert, a medical doctor for example.
Context
Articles, blogs, tweets, business supplements, all seem to be overflowing with articles promising to explain the difference between machine learning and AI.Much of that content is of poor quality, seemingly copied and pasted from the same sources. Some of it is actually misleading.
Despite this, it is worth recognising that there is a genuine need that attracts these promises of clarity.
Problem
That need is business leaders and policy makers trying to cut through the opaque terminology to better understand what they're reading, discussing and making decisions about.To make matters worse, the terminology around AI has been used in different ways through history, and this continues today even amongst experts.
Solution
We could encyclopaedically list out all the many various uses of the terminology and argue about their correctness.A more useful approach is to consider which definitions are going to be most useful to business leaders who are trying explore the opportunities of AI, clarify its operational impact, and understand the risks they need to own and manage.
In that context the following definitions are useful:
- Machine learning (ML) is the use of sophisticated techniques to learn patterns and insights from data.
- AI, artificial intelligence, refers to automated decision making that uses the insights from the previous machine learning step.
Not everything should be labelled machine learning or AI, even if marketing teams are furiously trying to attach the labels to everything they can.
A simple test is whether the AI has achieved a level of competence comparable to a human expert in a given task. If the machine learning process requires a human expert to encode lots of his or her knowledge into the process, then it is not really machine learning, it is simply a translation of that knowledge into computer instructions.
Illustration
Let's bring those definitions to life with an example. Imagine we want to separate millions of photos of cats and dogs.
If a person was to do this task it would be slow, laborious, and because people get tired, there is an increasing risk of error too.
We could try to get a computer to do this sorting. A computer doesn't get bored, nor tired, and modern computers are very fast and can crunch through a million photos within minutes.
We could teach a computer to tell the difference between a cat and a dog by writing down detailed rules with the help of an expert, a zoologist in this case.
This might work. We can describe rules that ensure the animals have two eyes, four legs and a tail. To separate cats from dogs, our rules might say that cats have different proportions for head and leg size to dogs. This is more difficult to get right.
In fact, even writing rules to find four legs in a photo is difficult. Some legs might be hidden, or the angle might make it hard to distinguish them.
This approach of trying to write very detailed rules doesn't generally work well. The the rules themselves can be difficult to get right, the process of preparing them is hard, and there are many scenarios where we don't know the rules in sufficient detail.
An alternative approach is to see if we can get a computer to work out the rules which separate cats and dogs by itself. This is machine learning.
The computer learns by being shown lots and lots of labelled cat and dog photos. Each example tweaks the computer's understanding a little bit. Initially the computer's understanding is non-existent, but after many training examples, the computer's understanding of what makes a photo a cat or a dog has improved.
You can see parallels between this gradual training process and pavlovian learning.
Once we've trained the computer, we can start using it to classify photos that it hasn't seen before. This automated decision making, based on a previous machine learning step, is called artificial intelligence, or AI.
In practice, not all attempts at machine learning work.
And even when they do, we need to carefully monitor how well the computer performs on unseen data. It may be that the accuracy can be improved with more training, or it may be that our particular architecture will never reach the desired accuracy.
Thoughts
Today, we're very good at teaching computers to recognise objects in photos. In fact, in some scenarios, computers can identify them more accurately than humans. For example, computers can be taught to separate certain kinds of malignant tumours from benign ones.This is impressive enough.
But the real power of machine learning and AI is demonstrated when this can be done more accurately than a human expert, a medical doctor for example.