A deep dive into deep learning

Romain Lerallut, head of Criteo’s AI Lab, looks at what advertisers really need to know about deep learning Deep learning might just be the pocket knife we’ve been waiting for, but for complex scenarios like digital advertising you still need purpose-buil

Criteo’s Romain Lerallut explains what deep learning can do for the industry, and where its limitations lie at the moment.

When we encounter a new technology, we tend to call it a revolution. While some innovations do change industries, the reality is that most don’t make much of an impact. The advertising industry is confronted with these “game changers” nearly every day, so it’s important to understand what is really going to transform your business and what’s just going to cause unnecessary headaches. Intricate and often misunderstood technologies such as deep learning are especially difficult to assess.

Deep learning is the newest member of the AI family and it is being called the solution to complex prediction, relevancy and banner blindness. When reading articles about it in the news, one can get the feeling it’s a turn-key, one-size-fits-all solution to every problem that the digital advertising industry is facing.

An engineer might think: “Wow, how could I have missed an entire development cycle, during which this absolutely crucial advertising technology was created?” Well, there are two misconceptions in that thought process. First, this isn’t a new topic. Researchers have spent more than 20 years on the subject of deep learning and have made significant progress in many domains, including image recognition and sound processing. Second, while the technology isn’t new, we’re just starting to scratch the potential of deep learning in advertising, so you haven’t missed anything.

WHAT IS DEEP LEARNING?

It’s important to start with a clear understanding of machine learning, because deep learning is just a subset of that discipline. There are three important areas to understand:

Supervised machine learning starts with a human definition, which the algorithm learns to recognise and categorise. A well-known example is a spam filter, which spots pre- defined features, like “free drugs” or “you’re the winner of a brand-new car”.

Unsupervised machine learning starts with uncategorised data, which the algorithm breaks down into groups by the interpretable clusters of patterns it recognises. From there, it’s up to a person to interpret what the data means.

Deep learning, in theory, joins the best of both worlds. Not only do you not need to define the patterns it’s looking for, but you don’t need to explain what those patterns mean either. Deep learning can recognise that a picture of a dog shows a dog, without a person feeding its features into the machine upfront (supervised) or analysing groups of interpretable features like legs and tails to find a meaningful data set (unsupervised).

THE ALL-PURPOSE TOOL

The way we talk about deep learning often reminds me of TV commercials for all-purpose tools. They promise to solve all the problems in the do-ityourself community. Take the Swiss Army Knife, the prime example for all-purpose tools. You can open a bottle, fix a bicycle chain, insert a missing screw and much more. An all-purpose tool might be the best tool you own, but it’s not the only tool you need. When the problem is more complex, like building an entire wardrobe cabinet, you need more robust tools that are purpose-built for the task.

Deep learning might just be the pocket knife we’ve been waiting for, but for complex scenarios like digital advertising, it’s not ready to make a significant impact. For that, you still need purposebuilt tools.

MAKING DEEP LEARNING A DIGITAL ADVERTISING TOOL

Implementing a deep-learning architecture in digital advertising would mean processing a lot more data than in other applications, such as image recognition, and doing so in real-time. Our advertising exchanges are driven by programmatic buying technologies that are under much stricter latency constraints than other use cases of deep learning – single- digit milliseconds at most.

Right now, the immense computing power for a true deep-learning architecture is not easily feasible. Instead, we’re seeing deep-learning models used to prepare in advance the information required for real-time decision-making. The difference is that deep learning is used to extract and process long-term user-centric information, instead of the more effective end-to-end approach that enabled critical gains in less constrained applications. But putting all that aside, is it any better than traditional models? Not necessarily.

Deep learning uses layers that process large amounts of raw data. The lowest layers figure out core features and higher layers then put it all together. Unfortunately, for deep learning models in digital advertising, the raw data is not as low-level as, say, pixels in an image. A deep model may be able to reciprocate hand-crafted features, like the time since a user’s last visit on a retailer’s website, but is that the best use of processing resources?

DIFFERENT NOT BETTER

It’s not that deep learning is better than traditional machine learning or vice-versa, but you need to consider your objective. Deep learning will absolutely affect advertising performance in the future, but only in the context of the whole machine-learning spectrum.

The important thing is recognising that every tool has its use. A screwdriver is not better than a hammer, because they’re tools built for different tasks. To understand how deep learning can help your business, start by following the scientific method and running experiments on your own data and your own KPIs. Measure which solution benefits you more, regardless of its internal implementation or the hype you read.