As the ecosystem gradually transitions away from text-based search and focuses its attention on image- and voice-based search, algorithms will need to exist for the machines to learn visual and audio cues that identify products, brands and intent.
Announced in May, Google's AutoML project is an approach that automates the design of machine-learning models. In effect it aims to apply machine-learning techniques to the process of designing machine-learning systems. With results confined to small datasets, AutoML is meant to design neural networks that would accomplish different tasks, or achieve better performance at the same tasks, than human-designed neural networks.
Yitaek Hwang, the director of R&D at Leverege, has said that AutoML will lessen dependency on intuition by iteratively trying out an algorithm, scoring its performance and choosing and refining other models.
Einstein Vision by Salesforce operates on a similar concept, helping Coca-Cola install smart sensors and cams inside coolers to recognize when a cooler is in need of replenishment, connected to the supply chain operation of the retailer and supplier.
In a blog post last week, researchers at Google revealed that AutoML can perform on larger and more challenging datasets to identify location, objects, and the underlying story that is driving the human experience. In an example cited, AutoML was able to identify the number of people in a picture at a beach, the number of people in and out of the water, and the number of kites flying in the sky over the ocean.
Why this matters to advertisers and agencies
"AutoML matters for advertisers and agencies because they don't have big-data and machine-learning experts on staff," said Augustine Fou, an ad fraud researcher and former advertising professor at NYU. "As more digital advertising goes programmatic, this could be a 'front lines' defense for the obvious forms of ad fraud." He hastened to add that caution is warranted because AutoML alone will not be able to root out more advanced forms of ad fraud.
Advertisers and agencies currently compete for target-segment eyeballs based on text-based search terms, aiming to index the relationship between their offering with the keyword being bid for. When AutoML reaches mainstream adoption, advertisers and agencies will be scrambling to create an association between images and voice with their unique offering.
For market leaders, the time is now to invest in image- and voice-based search in order to feed existing and emerging search platforms, thereby indexing an association between a visual or verbal cue and the product or service being offered.
Also worth noting
At this point, Google has not perfected its own capabilities around artificial intelligence and machine learning relating to video and audio identification and content sentiment.
The most prominent example of this is the YouTube algorithm, which has been surrounded with controversy. The YouTube creator community complains that it wrongly flags valid videos as copyright infringement and 'demonetises' them (removes the creator's ability to make ad revenue). For example, in the past week, influential YouTube creators were hit with demonetization notices for making reviews around the iPhone X and a new Nintendo game.
If AutoML results in an enhanced ability to create and train machine-learning models, it could help Google solve its own issues as well as opening new vistas for marketers.