When it comes to automation, we tend to think in extremes: either an all-encompassing generic AI that can do most things a human can, or a simple-minded one-task wonder.
Generic AI models are trained on large data sets that attempt to mimic the kind of inputs humans receive in learning about the world around them. These are large data sets spanning across domains, intended to equip the AI to handle all kinds of diverse and unpredictable use cases. This makes it quite powerful, but there’s a catch. Businesses such as retail and distribution typically have use cases that are specific to these industries. It’s possible that the generic AI is familiar with data that is relevant to your business. It’s also possible that it’s not. So whether it actually fits your particular requirement at the time of deployment can be hit or miss.
Take for example, a store that lets their customers order items via instant messaging. A customer sends a photograph of the product they want. The image is blurry. No problem, AI can fix that. AI can also extract the text from the image and figure out the name of the product. So far, so good.
But the customer has some more specifications. She wants a zero-calorie version of the same product, and in a different flavor, something citrusy. She wants you to get all the citrus variants up and show her some reviews so she can decide. Oh, and by the way, she has a peanut allergy.
A generic AI may or may not understand the diverse needs of this customer. Maybe it will learn after some customers express dissatisfaction when it brings up banana flavor. Hopefully this will happen soon enough so that the AI learns the meaning of ‘citrusy’ before the business starts losing customers. But until then it’s a hit or miss. If a business is prepared to spend a lot of time and data on training a generic model, it might just work someday. But most corporations don’t have the patience or money to invest in this kind of thing, and many such implementations end up failing.
So what could work? Look at the problem again. All of this customer’s asks are a set of specific use cases. You could train a model to detect blurriness in an image. You could train another to extract text from an image. And one to extract specific text related to calorie count. By smart use of attribute values, you could even line up all the citrus flavors for the customer to choose from, none of them containing peanuts. As for reviews, yes, a model could bring those up as well, but only after – the internet being what it is – checking for offensive content.
All of these tasks are done by different models. Each trained with specific use cases. Each trained for narrow focus and precision. Each doing one job, and doing it well. They are faster to deploy than one big generic model that has learnt to do most of the things you probably have no use for. They also work with greater accuracy since they are trained for a specific business.
But aren’t all these models too narrow on their own? Can they provide a seamless experience to the user?
Yes, if you get them to work together. And that, folks, is what we call a colleXion.
Define your needs, pick the models you need, integrate via APIs, and you’re ready to deploy!
ColleXion – your one stop marketplace for all things AI. Watch this space for more!