Machine learning versus deep learning

How do they differ?

Before we look more closely at convolutional neural networks, and the layers and the other components they are made out of, let’s just recap the fundamentals of what we understand by the term machine learning, and the subtle ways in which it differs from deep learning.

If the concepts in this activity are new to you, and you are having difficulty following them, we again recommend the preceding course in this series Machine Learning for Image Data.

Machine learning

In machine learning, the aim is generally to take a set of data and extract some underlying information or structure from it. In supervised learning this might be to classify pieces of data into known categories, while in unsupervised learning this might be to split the data into discrete clusters.

It usually follows the steps listed in the diagram above of:

• pre-processing and feature extraction
• use of any one of a large set of machine learning algorithms (e.g. decision trees, linear regression, K-means clustering, multi-layer perceptron etc.)
• trained model as an output, that can then be used to make further predictions.

What features are extracted, and algorithm used depends on the desired output, and selecting them can require a lot of knowledge and expertise.

Deep learning

In deep learning the aims are often very similar or identical to those of machine learning. The aim is still to take a set of data and extract some underlying information or structure from it.

However, there are a couple of subtle differences as shown in the diagram and remarks below.

In deep learning:

• data may still need to be pre-processed in some way, but feature extraction is not commonly done by the user
• in deep learning the algorithm used is almost always a form of convolutional neural network (CNN)
• the output after training is conceptually similar to machine learning, it is a model that can then be used to make further predictions
• commonly, deep learning networks are used for classification tasks, such as image classification, but other types of tasks including regression, image segmentation, and object detection can also be performed.

In the next week of the course we will discuss Convolutional Neural Networks in depth, beginning with a tour of the building blocks of CNNs, by looking at the different types of layers, and the particular kind of data structures we use to pass data between layers, known as tensors.

Before that though, we will finish the first week by introducing the computer software tools we will use throughout the course.

Images (c) The University of Nottingham