Generic AI: Powerful But Not Useful
Few industries are currently growing and changing as quickly as the AI industry, which seems to produce a new revolutionary idea for some type of business every few weeks. Although some of these innovations end up taking off and creating massive profits, new jobs, and global benefits, like smart home technologies, others, like self-driving cars, seem to stall out or encounter problems that prevent them from achieving the success their creators had hoped for. It is difficult to predict which new forms and applications of AI will be successful, but those who have been watching the industry long enough can make decent guesses.
This is why many AI experts have been closely watching the development of “generic” AI solutions like Google Vision. Generic AI systems are typically backed by massive technological power (which is why Google is a great example), making them seem especially promising. However, by their very nature, generic AI are not “narrow”: they are trained on broad swathes of often public data and are incapable of being extremely specific in their analysis. For example, Google Vision purports to be able to analyze almost any image, picking out text, objects, and content that users might want to filter out of a search or a display. But although this could be useful to a user dealing with millions or billions or images, it does not present a solution for a company that wants to detect the precise color of a ceiling fan image, or find images that contain a particular brand of sneaker.
To put it simply, generic AI are not trained on industry-specific datasets, and so they are unable to provide relevant information to a particular industry. Their use cases, while large in scale, are very small in scope. Most businesses have no use for image detection AI that can only highlight portions of text. They need image software that can detect specific patterns, parse technical specifications, list brand names, and more. Generic AI is powerful, but of limited use.
Why We Need Vertical-Specific AI
The simplest way to make generic AI work for industry-specific (and even business-specific) use cases is to employ the process of “transfer learning.” Transfer learning means training the generic AI on proprietary datasets, rather than the publicly available generic datasets that generic AI already relies upon. In this way, a single generic AI could be trained by a paint distributor to identify specific paint names in images, by a grocery distributor to identify specific beer brands in images, and by a women’s clothing distributor to identify specific fabric patterns in images. The same generic AI could become an almost infinite number of vertical-specific solutions for businesses in need, providing the kind of specificity and better-than-human accuracy that businesses seek when they invest in artificial intelligence.
Something even better follows from the idea of transfer learning for various types of generic AI: a huge AI marketplace of vertical-specific solutions that could be directly applied by the consumer. This would enable businesses without data science teams and huge training budgets to take advantage of artificial intelligence in ways usually available only to big corporations. We could create a massive library of AI solutions just as we now have huge libraries of smartphone apps.
Although this approach would require businesses to train generic AI, this would not be very difficult for them since most companies already have the necessary data at their disposal. In the case of retailers and distributors, this data often comes in the form of their product catalogs. Generic AI from the marketplace would be ready to face specific challenges, and after training, it could even face unique ones.
The challenge is the creation of the AI itself. As previously discussed, access to data science teams is the barrier many businesses face when it comes to taking advantage of AI automation. These businesses cannot afford to create the AI they need from scratch, so who would supply generic AI solutions to the proposed marketplace?
Just as CrowdANALYTIX has used crowdsourcing to create AI solutions for our customers, this same approach could be applied to the creation of generic AI for a large marketplace. Members of our data science community could provide AI they have created, and businesses could train and implement it themselves. We would see more companies able to access the AI they need for their digital transformations and growth without having to budget for internal data science teams.
Although generic AI like Google Vision is alluring, these generic solutions require more steps to be made problem-specific and therefore helpful to the businesses that need them. Crowdsourcing an AI marketplace could be the answer to this problem.