There are various ways machines are trained using machine learning. This section explores the three types of Machine Learning (ML) and also explores decision trees.

Supervised Learning

• As the name suggests the algorithm is being supervised by an expert who labels the data. The algorithm is provided with data and their labels then it learns about the data.
• Most supervised learning algorithms use a combination of human labelling and decision trees (explored further below) to learn from the data.
  • For example, a human supervisor might give the algorithm a number of photos of apples and oranges. They then tell the algorithm whether the photo is an apple or an orange. 

Unsupervised Learning

• As opposed to supervised learning, the data is not provided with labels. Instead the machine algorithm is told that there are two different fruits in the data and it attempts to determine its own patterns and trends.
• This type of training is usually performed when labels aren’t available, or when there are massive amounts of data that are too difficult to label or when the patterns and trends are unknown.
  • For example, a person might give the algorithm a number of photos of apples and oranges. They then let the algorithm look for its own established patterns. A human supervisor checks the final outputs and then adjusts the training algorithm to achieve better efficiency or accuracy.

Deep Learning (DL)

• DL is the most complex type of machine learning and uses neural networks. When we think about “AI coming to life” we are thinking about deep learning.
• Neural networks are very complex, open-ended, decision trees that are modelled after the way neurons work in the human brain.
• DL uses something called reinforcement learning where the algorithm is asked to create its own assumptions about data and then apply those assumptions to a problem. This approach is very similar to how we learn about desirable and undesirable behaviour as children.
  • If the approach is successful, the algorithm optimizes itself based on its conclusions and any provided feedback.
  • If the approach is unsuccessful, the algorithm will adjust and try again, learning from its mistakes each time.

This brings us to decision trees. Decision trees are intuitive algorithms that mimic how humans make decisions. In the example below, the decision tree asks questions and based on the answer, it continues down the tree to arrive at a final conclusion. 

Adapted from: https://www.edureka.co/blog/machine-learning-tutorial/.

While this is a very simple example, machine learning algorithms can have hundreds of interconnected decision trees to help them learn from data. A decision tree is made up of three components:

• A node is a question being asked 
• A branch connects nodes together
• A leaf is the end point of a tree where a decision is made

A new branch is created from a node if more or new information is gained by following the branch.

Additional Activity: Grow a Decision Tree

Using your activity workbook, create your own decision tree thinking about the following:

• What decision do you want to make? (Creating a recipe, preparing for the weather, deciding which route to take to work, etc.)
• Feel free to add decisions to the recipe steps you created earlier!
• Who is the intended audience/user? 
• What data would inform this decision tree? 
• How might you introduce decision trees to your participants?