There is a lot of excitement and some confusion across the ad industry around machine learning, and for good reason.
The availability of cheap storage and processing has made sophisticated machine learning available to a much wider range of industries than what was available even five years ago. The media business has seen machine-learning solutions find homes in a wide variety of applications, from predicting how likely a user will click on an ad to classifying users in lookalike models and optimizing campaign delivery.
This wider use has brought increased attention to various components of machine learning, and that's where the confusion arises. Perhaps no piece of the machine-learning toolkit has grown in popularity quite like artificial intelligence (AI), which has become so widely discussed that many use the terms interchangeably, as evidenced by recent vendor surveys and press coverage.
The use of the term AI to encapsulate a broad variety of algorithms, statistical and otherwise, creates even more confusion. Vendor marketing teams have flipped the nomenclature and made machine learning a subset of AI because selling artificial intelligence is a more compelling narrative. With the way the marketing and trade press define AI, I could make an argument that any sufficiently complicated Excel spreadsheet would be classified as AI.
Treating the two terms as interchangeable or treating machine learning as a part of the AI landscape only complicates things. To better grasp machine learning and AI, it's important to unpack their histories and how they are used in the media business today.
Machine Learning: A Primer
Machine learning is useful when humans are forecasting, classifying or optimizing something that is really complicated, with many factors to be considered. People are limited by their cognitive abilities and time available, but machines can, with varying degrees of cleverness, take many more factors into account when trying to solve a problem.
In the context of direct marketing, marketers would historically buy lists from magazine publishers and send mail to those on the lists in the hope that subscriptions to magazines like Car & Driver matched with people's interests and possibly influenced their purchasing behavior. Media planners made these list-purchasing decisions based on their understanding of the magazine's audience, the target market and the advertiser's product.
In the 1970s, it became practical to process data for every US household, and machine-learning algorithms made the decisions about which homes received mail offers. Advertisers in a vertical like auto could now use data about households to predict the likelihood that someone in the home would be interested in a new car. The mail pieces were now directed to homes that were more likely to buy, rather than those who only showed an interest, leading to reduced costs and more sales generated.
The same kind of thing is done in digital advertising. Agencies and demand-side platforms collect data on purchasers of an advertiser's product and build propensity machine-learning models that assign a probability to every user based on how likely they are to buy the advertiser's product. Once marketers know a user's likelihood to buy, they can intelligently determine what they are willing to pay to reach that user on an ad exchange. The more likely users are to buy, the more an advertiser should be willing to pay. Performance advertising is built on the back of propensity models. All of this is machine learning.
So, why are advertisers just hearing about machine learning now? In the past, machine learning tended to be applied in areas where the value was very high, such as credit scoring or stock trading. Data scientists have had very good techniques and algorithms for a long while, with much of what is in use today initially conceived before the 1960s. Until recently, processing power and, equally importantly, storage, was expensive and slow, making use of machine learning impractical in many situations.
Over the past 10 years, three dynamics have changed the cost-to-value equation.
First, the free, open-source community has created platforms, such as Hadoop and Spark, that can process data very efficiently. Second, the cloud providers, including Amazon Web Services and Microsoft, have created infrastructure that can host these platforms and let them scale with demand. Third, good, pre-packaged and often free machine-learning software is available from a variety of sources. These three factors make machine learning a more practical approach to problems that were previously solved by brain power or trial and error.
There are several hundred techniques that can be thought of as machine learning. Think of it as a class of techniques, like arithmetic, with its types of operations: addition, subtraction, etc. Machine learning is the same. Typically, it is separated into techniques that have either a specific thing that they are looking to predict, which is called supervised machine learning, and classifying things, called unsupervised machine learning.
This brings us to AI. While I don't expect marketers to change their messaging, I would suggest that it is more appropriate to treat AI as a subset of machine learning, or as a separate discipline. In the same way that statistics is defined by its techniques, AI is a group of computing techniques and not a particular problem being solved, such as natural language processing or image recognition.
AI refers to a very limited subset of methodological techniques, including neural networks or expert systems, that mimic cognitive systems of human intelligence. The technique is what defines it as AI, not the problem being solved. If you look on the machine learning
Wikipedia page, the techniques used for typical AI applications are a subset of machine learning.
In media, there are many applications for AI methods, mostly focused on understanding language and pictures. The voice recognition behind Siri and Alexa is AI-based. Google's Search Safe, which identifies pictures and web pages that may not be appropriate for children, is routed in deep-learning models. The techniques that power all of these applications are machine learning-based.
The tricky part of using AI is that it requires the data scientist to "train" the model, and deep-learning models require a huge amount of known data to train against.
Say you want to train a deep-learning model to identify faces in pictures. You need to supply the model with hundreds of thousands of pictures where the faces are identified. In this way, the model learns what a face "looks" like. In order to build the model, you need to create the "training set." Creating the training set means that someone needs to indicate in those pictures where the faces are, which can take many thousands of hours of labor to build. The lack of "off-the-shelf" training data sets meant that it was very hard to build AI applications, but these data sets are becoming more and more available.
As machine learning becomes an even larger part of the digital media landscape, brands will have to continue to educate themselves on the nuances and applications of machine learning. Those who continue to misuse AI when they mean machine learning make it hard to communicate with practitioners and disconnect the advertising industry's use of the term AI from more common usage across industries and vendors.
The proliferation of the term AI as a proxy for a complicated set of algorithms broadens the meaning to make it useless. Going forward, the ad industry should use the term AI to refer to a set of specific techniques that are used in building solutions, such as neural networks or expert systems, and keep the marketing speak to a minimum.