The algorithms that deliver these experiences require not just data, but trained data, and refining data involves time and investment. As an investor in AI solutions, the first questions I typically ask are “Where do you get your data?” and “How do you train your data?”
It may come as a surprise to some marketers that when they decide to use AI-based technology, they must first be sure their data is trained to deliver those experiences. This can involve some heavy lifting, which often includes manual tagging. It also requires access to large data pools to continually refine the accuracy of predicted outcomes.
It’s no coincidence that the companies with some of the most mature and successful AI implementations are those with huge data sets (Google, Facebook, Amazon, Netflix and the like), which gives them a substantial advantage when it comes to training data, as these companies are all able to access and leverage massive activity from their users to improve the accuracy of their algorithms.
This kind of data training is referred to as having a “human in the loop.” Think about Facebook’s facial recognition technology that identifies friends’ faces for tagging. The algorithm has been able to accurately learn to recognize faces through the billions of photos tagged manually by users. The algorithms of other large players are trained similarly through user behaviors, which refine the accuracy of predicting all kinds of outcomes.
We are entering an era of data democratization. Those with less access to big data pools have some options: They can buy access to trained data, or walk the longer road of training their own data sets.
A typical approach is to employ teams (sometimes hundreds of people) to manually check and tag huge volumes of images.
There are now services sprouting up to provide assistance with data training. One such company is Austin-based Alegion, which provides “supervisory” training sets: data sets pre-trained to specific categories, so an algorithm can have a head start in recognizing the most commonly used data before moving on to a client’s proprietary data, which will likely need to be trained.
Alegion also provides “supervised data training,” drawing on virtual teams of data category specialists, mostly living in countries with more affordable rates, such as India, or Ukraine -- sort of like a Mechanical Turk for tagging data.
For some categories like fashion or furniture, where much imagery is similar across providers, there are pre-trained algorithms e-commerce companies can use for visual targeting. For example, there’s GrokStyle, which serves the furniture retail space. When a shopper clicks on a white couch, for instance, GrokStyle tees up the retailer’s inventory of other white couches to check out, which increases the chance of a purchase.
While some customized data training may be required of a retailer to make sure images are paired correctly, most standard images fit with the trained data provided, and a furniture retailer can be deploying the technology and reaping the rewards in short order.
Sean Bell, co-founder and CEO of GrokStyle, notes the success of this business concept: “The more broad the training set the better, and the furniture category is broad enough to provide millions of images across the web. We’ve been training our model since 2015.”
If you can’t find big data sets to mirror your needs, you might just have to create them, and companies like Neuromation (“where androids dream of electric sheep”) are doing just that by creating “synthetic learning environments”: computer-generated simulations used to train algorithms.
As an example, if you want to train an algorithm to recognize a video of a girl doing a cartwheel, you likely wouldn’t have access to the volume of imagery needed for the task. But Neuromation can generate thousands of simulations of girls cartwheeling, using creative techniques similar to those used in online games.
This is a big step that puts smaller players on the same level as video data king Google, which has access to an unending supply of video footage.
These are just a few examples of how marketers can better prepare data for AI algorithms -- and access the new oil of trained data.