Understanding Artificial Intelligence, Machine Learning and Data

This piece is a high-level explanation of the relationship between Artificial Intelligence, Machine Learning and Data.

“AI is akin to building a rocket ship. You need a huge engine and a lot of fuel. The rocket engine is the learning algorithms but the fuel is the huge amounts of data we can feed to these algorithms.”

Andrew Ng adjunct professor of Computer Science at Stanford University

AI Rocket-01


For thousands of years, we have designed tools to aid us in achieving goals. From the spear, to the wheel, to the plough - we have developed instruments to help us change the physical world around us with greater efficiency and effectiveness.

In the 20th century, something changed. In 1936, English mathematician Alan Turing published a paper which outlined the theoretical underpinnings for the computer. Years later, these information manipulating devices were brought into reality, which meant the scope of our inventions expanded into a completely new domain: thinking.

Once thought of as solely the realm of human beings, technology available to us today can complete information processing and decision making tasks that are not only comparable to human capabilities, but in many instances, far exceed them.

However, the ability to simulate actions that were previously thought to be the realm of humans alone does not mean that a computer is intelligent.

Your Calculator isn’t Smart

Definitions of artificial intelligence can vary. One of them is:

The capability of a machine to imitate intelligent human behavior.

If the ability to do arithmetic is considered an intelligent human behaviour, then by this definition, a calculator could be said to be a form of artificial intelligence.

Even though most basic calculators can perform feats of arithmetic that are impossible for humans to do, most people wouldn’t ascribe any level of intelligence to them.

So, a definition of AI that makes more sense is:

The capacity of a computer to perform operations analogous to learning and decision making in humans.

The key word here is decision. The utility AI provides is the ability to simulate and outsource our thinking and decision making. 

This then begs the question: how do we program AI to know what is the right or wrong choice?


Machine Learning

Like anything intelligent, AI needs to be taught. Machine Learning is how we teach AI. It's the area of computer science that allows computers to learn and act without an explicitly programmed function.

Some forms of AI are preprogrammed. While this may work in some cases, for many others it’s basically impossible, as the task is just far too complex. Instead of writing out predefined rules, what we can do instead is feed them a huge amount of data, and let them discover the rules over time.

Let’s think about how humans learn to categorise objects as an example. You have probably seen thousands of different types and shapes of chairs before. The picture below contains two objects which you and I would classify as chairs, though the form they have is substantially different to most conventional chairs.

angular chairs

Over time, repeated exposure and interaction with a number of objects that have the goal or action of ‘sitting’ associated with them has lead to you labeling that set of objects as ‘chairs’. If you were to come across a chair that you’ve never seen before in your life (like the ones above), you’d still know that it was a chair.

This is very similar to how we train certain AI systems. It's an example of one of two broad categories of machine learning called supervised machine learning.

Supervised Learning

In a supervised machine learning system, the machine is presented with data with a labelled outcome (just like in the chair example above).

The goal here is to build a model that can map inputs to an output. So in the case of an AI that recognises chairs, you’d feed it thousands, if not millions of pictures that are labelled as either ‘contains a chair’ or ‘does not contain a chair.’

Over time, the AI will begin to determine what sorts of patterns of pixels tend to be correlated with images of chairs. As you feed the machine more and more data, it gets better and better at figuring out what a chair is and isn’t.

This is an example of a categorical model, as the machine is classifying the data.

If the objective was a numeric value, the model would then be called a regression.

An example of this would be creating an AI that predicts the price of a home (output) based on a number of inputs, like square footage, number of rooms and postcode.

Because supervised learning requires labelled data, it's a lot more expensive and difficult to implement. Another category, unsupervised learning, doesn't have this problem.

Unsupervised Learning

Here, machines are provided with examples, but no specific outcomes or labels are provided. Instead, the machine tries to find interesting patterns in the data. There are a few ways a machine can view the data to try and ascertain patterns within. They are Clustering, Anomaly Detection and Association Discovery.

Let's use a list of transactions as an example.





In this instance, the machine would look for examples that are similar and group them together.

Here we have:

  • Two transactions that occurred on Wednesday
  • Both were authenticated by PIN
  • Both purchases were for Petrol
  • The amount spent is similar

Anomaly Detection


Here the machine would look for data that is a significant outlier. In this case, the $2500 expense is substantially higher than the others.

Association Discovery


Here the machine tries to determine associations between variables.

So when the customer is Adam and the account number is 1542, the zip code is usually 4067. Or, when the class of the purchase is 'food', the amount is normally close to $40.

Supervised vs Unsupervised Learning

A simple way of thinking about these two categories of machine learning is:

Supervised Learning is about trying to predict.

Some examples of questions that a supervised machine learning algorithm might be trying to answer is:

  • Is this cancer malignant?
  • Will this student fail?
  • Is there a chair in this photo?

Unsupervised learning is about trying to discover.

Questions that an unsupervised machine learning algorithm might be used to answer include:

  • Is this transaction unusual?
  • Which products tend to be purchased together?
  • Are these customers similar?

 Machine learning is enabling AI to do some amazing things already, and in many cases, it's out performing its fleshy overlords (us). Here are some examples where machine learning is being implemented:



Training competent AIs requires huge amounts of data. Humans find it pretty easy to do things like identify objects after only being exposed to them a few times, but then again, not only we do have the most complex machinery we know of in existence locked in our heads, but we are also the product of millions of years of evolution. Computer vision, on the other hand, has only been around for a few decades.

So instead of requiring a only a handful examples of an object category or concept as a human would, machine learning requires tremendous amounts of data.

While the digitisation of everything means that we're creating more data than ever, 90% of it is unstructured and therefore hard to process.

Andrew Ng, former head of Baidu AI Group and Google Brain, and adjunct professor of Computer Science at Stanford University has a useful analogy for building AI.

“AI is akin to building a rocket ship. You need a huge engine and a lot of fuel. The rocket engine is the learning algorithms but the fuel is the huge amounts of data we can feed to these algorithms.”


AI Rocket-01



Artificial Neural Networks and Deep Learning

So we know machine learning is an effective way of developing functional artificial intelligence, but what sorts of training systems are out there today?

While there are many ways to train these systems, it turns out that one of the most effective methods is to based on the architecture of the brain. They're called Artificial Neural Networks.

Here’s how they work:

Like organic neurons, artificial neurons transfer information to other neurons in the network. These artificial neurons are organised into layers, with each layer performing different sorts of transformations on their inputs, which are repeatedly fed forward until they reach the output layer. Generally, the deeper the layers, the more complex the task (this is where the phrase 'Deep Learning' comes from). Layers between the input and output layers are referred to as ‘hidden layers’, as their output is connected to the inputs of other neurons and is therefore not visible as a network output.

Below is an example of a neural network.

basic neural network


Let’s use facial recognition as an example of artificial neural network processing. The initial layers may be used to identify more specific features of the face like the pupil, nostril, teeth, etc. These specific features act as inputs for the next layer whose function would be to determine more abstract generalisations like ‘mouth, eye, nose’, etc. These results would be then fed forward until it reaches the output layer where the determination is made. For instance, yes, this is a picture of a face, or no, this is not a picture of a face.


Reading an AI's Mind

The hidden layers in advanced deep learning algorithms are like black boxes.

We understand the inputs that we give them and the outputs they spit out, but what’s happening in the layers hidden between them is a bit of a mystery.

As we mentioned at the start of the article, AI is a way for us to outsource our decision making capabilities. AI is predicted to pervade nearly all aspects of our lives, making medical, legal, and potentially military decisions. Given the importance of these determinations, the ability to understand the logical sequence of events that lead to why an AI makes certain choice seems like a necessity. The fact that we don’t have a clear understanding of what’s going on is a cause for serious concern.

These issues may manifest sooner than you might think, as driverless cars are already on the road. Picture this: A service like Uber or Lyft now have a fleet of self driving cars that you can use. During one trip, and for no apparent reason, one of these cars swerves randomly into oncoming traffic, causing a devastating accident. In an instance like this, it could be very difficult for the engineers to determine what exactly what went wrong.

This issue is only going to increase in importance as we develop greater AI capabilities. To put this issue into some context, the organisation 80,000 hours thinks that AI research is one of the best ways one can spend their time when trying to use their resources to do the most amount of good for humanity, as AI is a potential existential threat to humanity.

This makes sense when you consider the staggering abilities general artificial intelligence is predicted to have in the future. (More: Different types of AI)


AI will be everywhere

The ubiquity of the internet and smartphone prevalence around the world means that within the next decade or two, nearly every person on earth will have access to artificial intelligence with capabilities that may astound us.

Companies like Google, Amazon, Microsoft and others already provide cloud-based artificial intelligence solutions that are becoming increasingly more accessible, so even those who are not overly technical can tap into this cognitive potential.

Business-wide technology integration is a necessity in the 21st century, and making use of the decision-making capabilities of AI is something that will separate the winners from the losers in all industries. The scale at which these technical developments can be implemented just cannot be matched by man-power alone. Those who don’t adapt will fail.

The ways in which our world will change due to the rapidly increasing intelligence of machines is impossible to know, but we are well on our way to seeing it happen.

Tags: Machine Learning, Deep Learning, Big Data, Original


Let's Chat

Want to digitally transform your business?
We'd love to hear from you. Get in touch today.

Contact Us