What is the first thing that comes to mind when someone says AI? Is it you interacting with your Alexa – asking her to set an alarm for the morning? Or is it the Terminator and Skynet?
For me personally I see AI in three different ways:
I grew up in the great era of Science fiction films. To me the future seemed to be an exciting but dangerous place, with super intelligent machines turning from docile aide to terrifying superior, often leading to disaster (think HAL in 2001 Space Oddity). I think of this world of Artificial General Intelligence (also known as strong AI) as future AI.
It was at University that I discovered the more theoretical side of AI. The study of cognition and how the brain works, determining what made things intelligent and sentient. This area of study has been going on for centuries, however, we now apply this thinking to machines – what makes a machine intelligent / sentient and whether it is even possible? To me, this is classical AI.
This is the most accessible area of AI, one in which millions of us interact with daily – smart assistants; the likes of Siri, Google Home and Alexa. Typically, these are examples of AI performing very specific and narrow tasks such as answering questions, or being used for image recognition such as detecting cats in pictures. For this reason, it is often referred to as narrow or weak AI, but I see this area as contemporary AI.
Now, each of us have probably had different experiences or held different opinions about these three areas of AI. It then comes as no surprise that defining exactly what AI is – and what it isn’t – is somewhat of a challenge. Research conducted by Sage suggests that 46% of UK consumers have ‘no idea what AI is all about’. This lack of clarity is reinforced by a wide range of terms that people often use interchangeably, such as automation, machine learning, and deep learning.
When approaching AI with more of a technology lens, Gartner’s latest emerging technology hype cycle (2018) illustrates that AI is right at the very top of the ‘peak of inflated expectations’. This not only means that the expectation of AI has never been greater, but that these expectations are likely to be out of line with the reality of what is really possible today.
Obviously, this is an area that many people have an opinion on. You only need to google ‘What AI is’ to see the plethora of articles written on the subject, specifically how it relates to Machine Learning and other terms.
It is now generally accepted how AI, Machine Learning (ML) and Deep Learning (DL) all relate to one another. However, automation is often excluded from these comparisons, yet automation actually serves as the grounding for it all.
AI is generally seen as the pre-requisite for any automation task. However, this is misleading. Whilst AI inevitably leads to automation, jumping straight to AI often ends up with a race to the most technically complex solution. In a field where there is already a lack of clarity, this can lead to over-engineered solutions, which either overcook the brief or don’t meet it at all.
To counter this, if we frame AI in the context of automation, then this helps us avoid diving straight down a technical rabbit hole. This also provides clarity on the different terms involved, showing how they are subsets of one another – this also explains why there is often confusion about the terms.
- Automation is turning a task typically performed manually by a human into a process carried out by technology. This can be something as simple as a calculator or a SUM formula in a spreadsheet through to something much more sophisticated using complex algorithms to create output(s) from given input(s).
- Artificial Intelligence (AI) is a field of Computer Science focused on understanding how humans perform various functions that we constitute as intelligence to create synthetic or Artificial Intelligence. In Stuart J. Russell and Peter Norvig’s book Artificial Intelligence, A Modern Approach they outline the standardised definition taken from Alan Turing’s famous Turing Test (1950). Originally, this involved posing written questions to a computer. If the responses to the questions were consistent with a typical human response, then the machine would be considered intelligent. This was then taken a step further with the Total Turing Test, which included the need for the computer to also physically interact with the human who was posing the questions.
This resulted in six different disciplines constituting AI, which are still relevant today:
- Natural Language Processing (being able to communicate using language)
- Knowledge Representation (being able to store and retain information)
- Automated Reasoning (being able to answer questions using the stored information it has)
- Machine Learning (being able to adapt to new situations and detect patterns)
- Computer Vision (being able to perceive objects)
- Robotics (being able to move and physically interact with objects)
- Machine Learning (ML) is a sub-field of AI where computers identify patterns in data to either understand hidden structures or make predictions from it. The reason it is called Machine Learning, and why it differs from statistical modelling, is that the computer doesn’t need to be explicitly programmed to do this. A human creates the Machine Learning process and selects the algorithm or model, then the computer explores the data to come up with the right calibration of the model itself. As new data is fed to the computer, it adapts the parameters of the model to give an optimal classification or prediction. This is often used in detecting different customer segments based on common attributes or to power churn or repeat purchase propensity scoring models.We have applied ML in a variety of ways. With audience creation, we have used ML to create more dynamic target audiences for digital advertising by scoring people’s propensity to purchase. We have also used ML to power our attribution modelling algorithms
- Deep Learning (DL) is a specific field of Machine Learning, which uses a family of algorithms called Deep Neural Networks. These algorithms are designed to mimic the structure of neurons in the brain (hence the name), where multiple layers of ‘neurons’ receive input values and provide an output value. They represent the most recent advancements in the field of Machine Learning, thanks to increasing quantities of big data being available, as well as advancements and accessibility of cloud computing. They are also responsible for a wide range of consumer services, such as predictive typing on smart phone keyboards, self-driving cars, as well as being able to automatically classify different Instagram images based on the content. They are often referred to as black box algorithms, as the way they work is hard for a human to interpret. The ability to predict an outcome in a given context can be strong, but it is not necessarily easy to understand why. Our own work with Deep Learning has helped with information compression to consolidate a wide range of data signals into a smaller set of features to feed into predictive audience targeting. Deep Learning was the only way to achieve this due to the scale of data signals involved. We have also used Deep Learning to understand what type of digital video content our audiences are consuming. Given the alternative of manually reviewing each video, this was the only viable approach!
So, now that we have established what AI is, we can ask ourselves the real question: why are we interested in AI?
Firstly, as marketers, we need to understand why we are interested in it and how we are going to use it. We are too often tempted by the promise of shiny new toys without first asking why and if we should be using them in a particular context. This is core to how we approach problems at Annalect, with curiosity being an integral part of how we work.
How Annalect has used AI
One of the areas where we have applied Machine Learning is in developing predictive scoring models that calculate the probability that someone will convert or purchase. Doing this allows us to create dynamic audiences of warm prospects for more targeted digital advertising.
When it comes to determining a specific model to use for the actual scoring, we test a variety of options, starting with the simplest and increasing with complexity, only if there is a net improvement in prediction accuracy.
Often a simple model like logistic regression can provide a good enough solution to a classification or prediction problem, but there are situations where more complex solutions are needed.
When processing thousands of different data signals to create different audience segments, Deep Learning can be an effective solution to compress this information into more manageable groups.
The more manageable groups allow us to then create a base set of more manageable segments akin to industry segmentations, like Mosaic or Acorn, but built from large scale digital data. Other approaches to compressing information can work in different contexts, such as working with survey data, but in this case a more complex solution worked better for us.
Finally, there are some problems where you simply have no choice but to use Deep Learning. A typical example of this is image recognition. When we wanted to understand what digital videos our target audiences were interested in we had to employ deep learning techniques. Doing it manually was simply not possible, as it would require an army of people to watch all the videos and record what they see.
We used Deep Learning to classify the different types of videos that people are watching. This allows us to better understand our audiences to develop more relevant creative messaging.
Ultimately, whenever you are approaching a project, the key is to start by being clear on why you are considering AI, and what the solution needs to look like. If your question is about understanding the ROI you are getting from your marketing investment to help with strategic budget planning, then Deep Learning may not be the right solution. Being black box in nature can make it hard to interpret the results when presenting to senior stakeholders. Conversely, if you are interested in understanding which groups of customers are likely to repeat purchase, then Deep Learning may be a good solution – however, always start with more simplistic Machine Learning first.
Doing this means you can get a baseline view of how accurate your algorithm is, and whether it is allocating people into segments or predicting sales. You can then build on this base level and test more complex approaches, whilst being clear on whether they are providing a sufficient gain in performance.
AI undoubtedly presents an exciting opportunity for marketing now and in the future. However, to really start leveraging its power always start by asking why: why do you need it and what outcomes do you want to get? Once you have done this, starting simple and increasing complexity as necessary, you will always end up with a better solution to your marketing challenges.
By George Maynard, Managing Director, Annalect Labs