What is Deep Learning? Here's Everything Marketers Need to Know

The machines are here.

You may have heard rumors about artificial intelligence (AI) potentially taking over our jobs. And the question is: Should you be concerned?

In my opinion, we should be excited.

AI -- especially “deep learning” technology -- brings new opportunities and innovation in the way digital marketing, sales, and customer support are handled.

But what is deep learning? How does it work? And how can it be applied to marketing and sales in your company?

What Is Deep Learning?

Deep learning is a discipline within AI that uses algorithms mimicking the human brain. Deep learning algorithms use neural networks to learn a certain task. Neural networks consist of interconnected neurons that process data in both the human brain and computers.

Neural Networks in Advertising

Let’s assume we are an online car dealership, and we want to use real-time bidding (RTB) as a mechanism to buy ad space for our product on other websites -- for retargeting purposes.

RTB is an automated process that takes place in a short time frame of under 100 milliseconds. When a user visits a website, an advertiser is alerted, and a series of actions determines whether or not that advertiser bids for an ad display. Have a look at the image below:

Source: Periscopix

In RTB, we use software to decide if we want to bid for a certain ad -- the software will make a decision by predicting how likely the website visitor is to buy one of our products. We call that "buying propensity."

In this instance, we'll use deep learning to make this prediction. That means our RTB software will use a neural network to predict the buying propensity.

The neural network inside our RTB software consists of neurons and the connections between them. The neural network on the above image has only a handful of neurons. In reality, a digital neural network has thousands -- or even millions -- of neurons and connections. 

In this scenario, we want to find out if a certain website visitor is likely to buy a car, and if we should pay for an ad to target her. The result will depend on the interests and actions of the website visitor.

To predict the buying propensity, we first choose several “features” that are key to defining this person’s digital behavior. In our example, those features will consists of which of the following four web pages were visited:

1. Pricing.
2. Car Configurator.
3. Specifications.
4. Financing.

Those features will influence the output of our neural network -- or, essentially, our conclusion. That output can have one of two values:

1. The website visitor is interested in the product, or “ready to buy.” Conclusion: We should display an ad.
2. The website visitor is not interested in the product, or "not ready." Conclusion: Do not show an ad.

How the Neural Network Functions

Let's have a closer look:

For each input, we use “0” or “1”.

“1” means the user has visited the webpage. The neurons in the middle will add the values of their connected neurons using weights -- or, more simply put, they define the importance of each visited webpage.

This process continues from left to right, until we reach the “output” neurons -- “ready to buy” or “not ready,” as per our earlier list.

The higher the value of the output, the higher the probability that this output is the correct one -- or the more accurately the network predicts the user’s behavior.

In this example, a website visitor looked at the Pricing and Car Configurator pages, but she skipped Specifications and Financing. Using the numerical system above, we get a “score” of 0.7, which means that there is a 70% chance this user is “ready to buy” our product.

So, if we look at our original formula, that score indicates the conclusion that we should buy the RTB ad placement.

Training of the Neural Network

Now that we know how a neural network functions, let's have a look at how to make sure our output neurons are calculated correctly, in order to make the right decision.

The challenge is to come up with the correct “weight” factors for all the connections inside the neural network, which is why it needs to be trained.

In this context, “training” means that we feed the neural network data from multiple website visitors -- things like visitor features (which web pages users have visited), as well as indicators of their eventual purchase decisions from us (which are labeled as "yes" or "no").

The neural network processes all these data, adjusting the weights of each neuron until the neural network makes appropriate calculations for each person within the training data. Once that step is complete is done, the weights are fixed, and the neural network is can more accurately predict the outcome for new website visitors.

The Future of Deep Learning

Democratization of AI

AI is quickly finding its way into marketing tools that we use every day. Take, for example, the AI-powered Chatbot builder by Motion.ai (part of HubSpot), which allows you to easily create and publish your own chatbot.

Another example is Dialogflow, a platform from Google that lets you build a chatbot for your company or service.

It certainly doesn't stop there. AI can assist with the setup of advertising campaigns, hyper-personalize emails, optimize lead scoring, categorize and escalate customer issues, and actually help you with anything that requires data processing or orchestration.

Deep learning can be applied in any area of digital marketing, provided that you have a sufficient amount of “training” data. The challenge is typically to extract data from your various marketing tools -- that's where data integration platforms like Blendr.io will be crucial in connecting your data silos when you start experimenting with deep learning and AI.

The Future: AI ... That Builds New AI

Google explains that the process of designing neural networks often takes a significant amount of time for development and experimentation, because all of the neural network layers have to be crafted by people. That's why Google invented AutoML: AI that can build new and better AI algorithms.

Imagine what that type of technology can bring to something like marketing automation, for example. The AI will be able to build additional, customized AI algorithms that will learn and automatically optimize nurturing campaigns, for example.

Though deep learning may sound complicated, it's a process that, much of the time, boils down to math. Neural networks “learn” in a manner similar to humans: by viewing many examples, and discovering the commonalities among them.

Once the neural network is trained, it can perform complex tasks and a certain level of reasoning. Deep learning and AI can be integrated into many aspects of digital marketing and sales automation. The machines are not coming -- they are already here.