If we want to have a meaningful conversation around anything connected with artificial intelligence (AI), we need to be speaking the same language. Some terms come to be considered as interchangeable, or as meaning the same thing when actually they don’t.
‘Machine learning’ and ‘deep learning’ are a case in point. Adding to the confusion is that both are statistical modelling techniques, both can be used for making predictions, one is a subset of the other and they are both subsets of AI.
Both can also be utilised to produce results from massive data sets, for example, but only one is as useful when there are less data points available.
So how do we best understand the difference?
As with most complex ideas, it pays to simplify and to illustrate with examples. Let’s go through each technological offering individually, then bring them back together at the end to discuss similarities and differences.
But first, let’s agree that AI is simply a catch-all term for anything that applies any form of human intelligence to machines. All of the different types of learning – supervised and unsupervised, machine and deep, rule-based, predictive and cognitive – are captured under the ‘AI’ term.
What is machine learning?
Machine learning is the more basic of the two learning methods as it involves the machine being taught the meanings of a specific set of data so that it can eventually develop specific and pre-defined outcomes on its own.
Algorithms, or sets of rules that the machine must follow in order to produce an outcome, tell the machine what it must look for and how it must interpret data.
An example might be the fact that a machine can be taught to recognise common types of malware, despite the malware constantly changing shape and form. The coding within various types of malware tends to share specific similarities, meaning machines can be taught to weed out the nefarious applications from massive pools of data. As the malware develops and evolves, the machine’s knowledge does, too.
So in machine learning, a machine can modify its data analysis parameters in order to continue to produce the defined outcome. The machine is smart and it learns, but only toward a certain goal or purpose.
A simpler example of machine learning is a machine being able to tell the difference between an orange and an apple. In order to achieve this, the machine might be fed data relating to weight, shape, skin texture and colour, until it can reliably differentiate one from the other.
This is why projects with less data are more fitted to machine learning. If there are less data points to be used as specific references – such as weight, colour, texture, etc. from the example above – then machine learning will likely do the job admirably.
And while machine learning can also be used for complex, data-intensive tasks, such as predicting a user’s taste in music, it is fairly basic compared to deep learning. Deep learning is, if you like, the evolution of machine learning.
Let’s explore what that means, now.
What is deep learning?
If a computer has been taught to do so, it will be able to make sense of new information via analysis and to assign the result of that analysis into a pre-set category – orange or apple, malware or safe download, dog or cat, etc.
But what if it hasn’t been asked to come up with a specific outcome or to identify defined items? What if it could recognise entirely new patterns and trends, some that are too complex to be noticed by the human brain?
This brings us into the field of deep learning, which can generally be defined as the utilisation of neural networks that are many layers deep in order to provide an interpretation of the data. In other words, deep learning involves the analysis of large sets of data to identify outcomes, patterns, results or interpretations that have not been previously defined.
In this way, deep learning comes much closer to imitating the way the human brain works and, just like the human brain, it can have surprising results.
In deep learning, algorithms are many layers deep, allowing the machine to come to its own conclusions.
So, for example, imagine you have a collection of images of aeroplanes and helicopters. The machine learning method of telling them apart would be to label each image as ‘aeroplane’ or ‘helicopter’, then feed the machine enough examples that it builds up its own database of data points until it can eventually tell the difference on its own.
The deep learning method, on the other hand, would be for the machine to analyse the specific features of each image and come up with categories. To produce accurate outcomes more data would be required than in the machine learning method. However, once complete, deep learning will likely produce not just a category for helicopter and a category for aeroplane, but also various categories for different sizes of vehicle, colours, wing shapes, numbers of windows, and lots more.
Differences and similarities
If we consider a neural network as a computer system modelled on human thinking and analysis patterns, machine learning involves a single or double layer. Machine learning is like the toddler, discovering differences between two colours by using their vision.
Deep learning, on the other hand, is many neural networks deep. It is coming closer to using the full power of analysis and, when tasked with analysing massive data sets, will often recognise what the human mind cannot because of the sheer scale of data.
Machine learning relies on structured data from which pre-defined outcomes will result. Deep learning typically requires large amounts of complex data and will produce outcomes of its own accord.
Both machine learning and deep learning offer massive power to their users and both will play an increasingly important role in business, innovation, politics, investment and more.
Find out more about that new world of opportunity with our Master of Data Science.