Exploring types of machine learning

There are many machine learning algorithms available to the creative AI artist today. These algorithms are the building blocks upon which the creative AI practitioner builds their machine learning system.

Supervised and unsupervised learning

However, in order to identify the correct algorithm to use, the artists must first learn to identify what problem they are trying to solve. There are two main types of machine learning techniques. These are supervised and unsupervised learning.

Put simply, if you have a huge amount of unlabelled data and want to organise it and discover patterns, you would use unsupervised learning. If you have labelled data and want to predict future trends based on this data or identify new objects based on previously labelled objects, you would use supervised learning.

Within supervised learning, the two main types of machine learning algorithm are called classification and regression. Again, which one you will choose depends on the problem you are trying to solve, and the type of output your task requires. If your output takes the form of discrete labels you would most likely want a classification algorithm. Alternatively, if you want a smooth stream of data as your output, then you most likely want a regression algorithm.

Let’s look at these examples in more detail.

Classification

Classification is the type of machine learning algorithm used to identify discrete types of data. The data can take any form, for example, photos.

If you wanted to be able to teach a machine to identify pictures of cats and dogs, you would use a classification algorithm. To train the machine, you would start by assembling a set of photos, labelling each photo as a cat or a dog. You call this set of photos your ‘training set’. You would then input this training set into the algorithm as many times as you like. The machine can then learn to associate the labels ‘cat’ and ‘dog’ with the appropriate photos. Once this training has finished you would then be able to show the machine a new photo that is not included in the training set. If the machine has been trained correctly, it will then hopefully be able to identify whether it is a cat or a dog in the photo.

This is called supervised training because the researcher is telling the machine what is contained in the photos by explicitly providing labels.

_{Image of classification vs regression}

Regression

If the data you are trying to analyse can not be approximated using discrete labels, then you will probably want to use a regression algorithm.

Regression is a type of statistical analysis used to identify trends in data and fit this data to a curve. If the data can be approximated in this way, it allows the algorithm to predict trends in data that isn’t in the training set. The data can be multidimensional. By that we mean that many parameters of the data can be measured. The regression algorithm will then represent the relationships between all these parameters. This allows the researcher to smoothly map parameters within a defined parameter space.

Generation

Certain types of neural networks have been used very successfully to actually generate media based on, but different from, the media used to train them. Recurrent Neural Networks (RNNs) are an example of this. They are basically traditional neural networks, but with a loop built into them so they can iteratively process data and learn from their previous state. A specific example of an RNN is the Long Short-Term Memory (LSTM) network. This type of network has been used to generate novel media such as text and music.

The MIMIC project is an initiative run by a consortium of UK universities that is leading the way in the use of machine learning for creative applications. They are developing an online resource where anyone can go and try their hand at creating machine learning systems right in the browser. You can find out more, and explore the project yourself, by visiting the links in the ‘See also’ section at the end of this step.

Have your say

Given the techniques discussed above, think about how they are being applied to work you have seen so far in this course.

• Can you find any examples of regression or classification being used in creative work that has not been shown already?
• Have a look at the websites of the artists we have met. Use their work as a clue to search for similar artists or work being made in the field.

Share your findings with your fellow learners, and discuss how you think data is being used in the examples you find.