ML-based solutions: theory VS real (business) life
June 22, 2021

Here is what actually happens when Machine Learning meets the market

It is quite usual to talk about Artificial Intelligence on a very technical and academic level (and we have already considered this approach in many previous posts). And it is also quite common to read posts and website pages in which companies tell the world that they deliver “ML-based solutions”. But is it actually true? Do you really know what happens when Machine Learning models need to concretely face market’s and customers’ needs

Previous blog posts

A business-centred point of view on Artificial Intelligence

As mentioned before, often AI is not taken into account from a business-centered point of view, even if many companies use “AI” as a magic word when they talk about their products. Actually this kind of technology needs to be approached in a specific way to really create effective solutions for the market.

Without going too deep in ML-Ops related issues, the image shown above is a crystal clear example of this field’s complexity. A Machine Learning model can be very accurate in distinguishing between pictures of wolves and pictures of huskies. But, what if its accuracy was based on the snow in the background? Would you rely on a system that is not based on the awareness about its sector of application and its starting needs? In that case, the problem is not taking into account the huge difference between correlation and causality, and it often happens that these two concepts are erroneously overlapped. But this issue becomes even bigger when Artificial Intelligence is used for real solutions, stepping out of the academia research.

Win the complexity challenge to reach effectiveness: Aptus.AI approach

At Aptus.AI we have clear the complexity of developing effective AI solutions for the real world, and not just functioning models. This is why we have created - and we follow - a tailored and tested approach, which has its roots in a deep knowledge of Machine Learning and mostly in a clear awareness of the difficulties that developing ML-based products presents.

The latter can substantially be divided into two macro-types, that are:

  • Product-related aspects
  • Technical aspects

Starting from the first kind of issues, we can identify three further levels of the problems depending on the product itself, namely:

  1. User expectations
    In respect to ML-based solutions, often users don’t know what to expect from the product, because they are not aware of what AI really does. Some examples? Google Home can tell you the age of any famous person in the world but never their age difference. People who don't know that the feature is based on search intent, maybe can find it really weird. Often users expect the wrong things from Artificial Intelligence tools. This is why at Aptus.AI we always build the development of our ML-based solutions in cooperation with our customers, in order to have all clear in mind what the system can do and what it cannot do.

  2. Unexpected behaviors
    Apart from users’ expectations, AI solutions can be very disappointing also for their behaviors. Even if they are very accurate, some of them can result actually surprising when they make mistakes. Let’s not forget that Machine Learning systems are fed with datasets, but we can say that they learn on their own. So even being 99% accurate, they are not working as code written by a human, so their errors are totally unpredictable. Not clear enough? No problem: just watch the video below and you will surely understand what we mean!

  1. Ethical issues
    Finally, there are also ethical aspects to be considered in respect to the development of AI products. This time we are referring mostly to common sense. These kinds of issues cannot be universally solved with definitions or formulas. Would you think that a software to identify criminals just from their face would be right? Well, it has been developed - Phrenology 2.0. Obviously, these aspects depend on what is the field of ML application and how it is specifically used. This is why at Aptus.AI we consider it essential also to dedicate the maximum attention to the product development and management, starting from the very first steps of our projects.

Some lines above we have talked about a “tailored and tested approach”, and it was not just an expression. At Aptus.AI we have implemented a specific workflow which goes beyond the normal steps of model design and offline evaluation used in the academia world to develop ML systems. And this is how we have created Daitomic, our AI compliance management platform addressed to the RegTech market. What are the further steps that we have added? Keep following us to discover them - starting from next post!

Be the first to try the AI tool that will revolutionize financial compliance