“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Today’s column is written by Santanu Kolay, senior vice president of engineering at Turn.
Artificial intelligence (AI) is one of the most-hyped topics in advertising right now.
At Cannes, Saatchi & Saatchi featured an AI-created film. This summer, IBM’s Watson rolled out AI-powered ads for The Weather Co. that answered consumer questions.
But the reality is that while there is a lot of smart tech being applied in the industry – deep learning, machine learning and algorithms – we’re still a way off from true advertising AI. The AI-in-ad-tech story sounds good but it’s still mostly fictional.
Some describe AI as the capacity for learning, reasoning and understanding. The famous Turing test has been touted as the key requirement for success, asking, “Does the machine show intelligence equal to that of a human?”
The ‘Ancient’ Machines
Machine intelligence and AI have intrigued researchers since the 1950s. Initial AI systems were mostly rule-based with rules created by people, and the AI systems behaved more like an information retrieval system. Later, the research community shifted toward machine learning, where the focus is on pattern recognition from vast amounts of data. Recently, there has been a surge in applying deep-learning techniques to augment traditional machine learning.
Deep learning is expected to produce significantly better results with pattern recognition problems, and it excels in domains where there is a lot of data and correlation (spatial or temporal) present in the data. That’s why the most successful deep-learning applications have been limited to image or speech recognition and natural language processing.
What’s Happening Now
Facebook, Google, Amazon and Microsoft are at the forefront of AI and are focusing on solving problems pertaining to the areas where deep learning provides the most benefit. They are all getting better at understanding spoken language and figuring out what consumers are asking for.
Amazon’s Alexa uses speech recognition to interact with users and learn their daily pattern of activities so it can make recommendations based on activity patterns, but it must understand each activity accurately. Each of these individual tasks are better performed with the accuracy achieved by deep learning, and only then can each step be connected through rules – either learned through algorithms or created by a human – to achieve the ultimate AI.
In advertising technology, we deal with a different set of problems. We have less data, and the correlation in user behavior is often not strong, so applying deep-learning methods to computational advertising problems is harder and, at a minimum, the resulting improvements from deep-learning methods are not significant.
Ad tech companies are more interested in delivering ROI to advertisers, requiring them to improve the accuracy of their customer behavior predictions. Much of this improvement can be achieved by tuning existing machine-learning algorithms and using selective deep-learning methods to augment traditional machine-learning methods. Examples of deep-learning methods showing promise include low-dimensional embedding of high-dimensional data and recursive neural networks for sequential event prediction.
Getting To True AI
AI is expensive. It requires complex algorithms, highly trained workers and specialized hardware that add to infrastructure costs. At the moment, the benefits of the AI available today don’t justify the expense in digital advertising, where margins are under pressure. This is why we’re largely seeing incremental improvements to existing algorithms in the ad tech space. We are still far away from the holy grail of AI: a fully automated system that requires no human intervention.
What then? As mentioned, there are things machines can’t seem to duplicate. It’s hard to see how machines could come up with a funny ad, for instance. What they could do, however, is create permutations of ads and then determine – based on human reaction – which were most effective. That’s the kind of intelligent combinatorial exploration we’ve seen in man-machine matchups in chess and Go.
Another possible use case of AI could be simplifying ad tech workflow, where the AI learns from human interactions with an ad tech platform and learns to perform some similar optimizations based on the training.
Decision-making draws on factors beyond data that include intuition, empathy and knowledge of human thinking. In other words, if you want to draw an emotional reaction from a human, you need another human to contribute to it. AI is coming to the advertising world but don’t hold your breath waiting for the rise of the machines.
Follow Turn (@TurnPlatform) and AdExchanger (@adexchanger).
