Chasing Black Swans
A further key advantage of synthetic data is the ability to model situations that are incredibly rare in the real-world. This means businesses do not have to wait for infrequent or one-off ‘black swan’ events but can train algorithms based on synthetic data.
“You can create synthetic data for everything, for any use case, which brings us to the most important advantage of synthetic data – its ability to provide training data for even the rarest occurrences that by their nature don’t have real coverage.”
Payment services firm American Express recently used this approach to train AI models for unique, or atypical, types of fraud, where insufficient real-world data existed.
“There are a lot of different kinds of patterns, the number of fraud patterns in real life is pretty big. Some fraud patterns happen more often than others, and some patterns are very rare,” Dmitry Efimov, head of Machine Learning at American Express, explains, “We started thinking, can we balance the presence of different fraud patterns?”
Using a type of AI algorithm, known as a generative adversarial network, American Express was able to create fake fraud data that was then used to train a second AI, all without the need for any personally identifiable information at any stage of the process.
While this pure approach can work for some well-defined problems, the majority of researchers today view synthetic data as a hybrid solution, where it is blended with some real-world data to improve accuracy and speed. AI startup Neuromation estimates that companies only need 50% of their original, authentic training data to finish the formal training of their algorithms.
SIGN UP TO OUR NEWSLETTER
We do not sell our lists, and you can easily unsubscribe if you so wish.
Fake data risks
Alongside the obvious benefits of synthetic data there are however sizeable risks. Most notably, the disconnect from real-world events can make it difficult or impossible to verify the provenance of data points. In the simple case this may just mean that the machine learning algorithm takes longer to achieve accurate results, but it also opens the door for malicious actors or deliberate bias.
Srividya Sridharan, Senior Research Director at advisory group Forester sees an emerging split between “the good, the bad, and the ugly of artificial data” with positive benefits coming from “synthetic data that allows users to create data sets for training AI” and serious risks from “fake data that does the opposite [and] perturbs training data to deliberately throw off AI.”
As firms face increased pressure to verify the source of the data used to train their AI models bias, they may face difficulties, made even more challenging by the increasing number of for-profit firms selling proprietary synthetic data sets.
Deb Raji, a technology fellow at the AI Now Institute explains that many synthetic data sets can actually exacerbate the biases, noting that while it can be useful for assessment and evaluation it can also be “dangerous and ultimately misleading” when it comes to training AI. Ian Coe, CEO of synthetic data startup Tonic.ai, reinforce this point, noting that getting quality, safe data to developers is “a gargantuan task in and of itself” and “it’s just the start”.
While Big Data will continue to dominate businesses for years to come it seems certain that the form that data takes will continue to evolve as privacy, cost and competition shape the market.