What advertisers really need to know about Deep Learning

By Colin Barnard, Commercial Director at Criteo, ANZ

Deep Learning is often perceived as the solution to almost all problems in digital advertising. As the youngest child of the AI cosmos, Deep Learning promises to increase relevance, improve prediction and reduce banner blindness.

Deep Learning isn’t a child anymore. Researchers have studied Deep Learning for more than 20 years with significant developments in image recognition and text and sound processing. But how do these new developments relate to digital advertising? Especially considering ads are implemented in real-time and utilise more complex data than standard bare pixels and sound frames.

The Swiss Army knife of digital advertising

You could say Deep Learning is regarded as the Swiss Army knife of digital advertising. The Swiss Army knife is a multi-purpose tool but to solve more complex scenarios, we often need to use a Swiss Army knife in conjunction with other tools or source the help of a handyman. 

This notion can be applied to Deep Learning. Deep Learning may be the most powerful subset of Machine Learning, perfectly suited to solve tasks like image recognition, but for digital advertising where large amounts of quality data and years of experience are required, this all-purpose tool shouldn’t be the only ace up a marketer’s sleeve.

Machine vs Deep Learning?

It’s important to understand how Deep Learning fits into the Machine Learning spectrum:

• Supervised Machine Learning teaches algorithms to view data and cluster them respectively to make predictions. An everyday example of this is the spam filter in our inbox. In digital advertising, Deep Learning can predict the likeliness of a user clicking on a banner for example. It involves defining the features to generate a label as an output, like “This mail is spam” or “This user has a Click Through Rate of X”.
• Unsupervised Machine Learning finds patterns in a large pool of data, analysing the calculations’ results to classify consumer behaviour. You don’t have to define a feature or label, instead the machine looks for interpretable clusters of patterns.
• Deep Learning is a subset of Machine Learning. Based on artificial neural networks, Deep Learning can be supervised, semi-supervised or unsupervised using high-speed machines that have computed, analysed and stored huge amounts of data in the past. But there’s a catch.

Should advertisers use Deep Learning? The answer is just a little.  

Mastering Machine Learning

The challenge with Deep Learning is that it requires an enormous amount of data to train such a complex system and for advertisers this means processing the data in real-time.