Artificial Intelligence needs engineers, not data specialists !

I just read an article mentioning that AI is both overhyped and underexploited, appearing as a paradox... It shows very well how things are turning for AI nowadays : too many announcements of (so-called) AI applications with poor results, and lack of AI usage in some fields where it could actually be useful. Where does it come from ?

As many technologies in the past, there is a major trend to put AI everywhere, and furthermore, to put deep learning (which is just a part of it, as there are many other possibilities). Mark Twain said once : "Every problem looks like a nail when the only tool you have is a hammer". And this is what we can see today : as deep learning proved to work well on several problems precedently hard to solve (playing Go game for instance), people try to apply it on everything. But DL (and generally data-based learning systems - mostly artificial neural networks) depend on the existence of data offering representativity (making sure every possibility is present in the data), accuracy (to cope with the actual level of precision required), and availability (data can be hard to obtain, or even too expensive) ; this is NOT a matter of size of the data set ("the more data, the better the result" is a myth).

In real-life applications it makes no sense to take an a priori technique and to try applying it to the problem : this is what we do in research (experimentation), and this is what the reference to Mark Twain stands for... We have seen in the past already this is the best way to fail : expert systems in the 80's, neural networks in the 90's, and other approaches in the "cognitive side" (case-based reasoning, multi-agent systems, etc). Today the hype is deep learning, just following the "big data wave". But should we forget everything that has been developed already on applied maths, from linear regression (simple but useful in about 80% of the problems) to bayesian models or complex data analysis ? Of course not : when we are dealing with industrial problems the point is effectiveness and efficiency, not hype !

What is my point ? Of course we need to go further with deep learning, when we have the data and when it is relevant (we should be able to decide that). But there are classes of problems where there are more simple ways to process. IBM Watson for example, is more using a cognitive approach to understand (big) data, in order to provide some reasoning. Neural networks are very good classifiers, but feeding them with raw data is not efficient at all, and the right approach is processing the data with "classical" techniques (data analysis, signal and image processing) in order to address the complexity of the problem where it is (non linearity, need for generalization, etc)... Using AI in industry is relevant in many classes of problems, when the right techniques are used on the right data : and this is an engineer job ! Data scientists are good at managing and processing huge sets of data, but they are not specially familiar with signal processing or automation processes that leads to integration to industry...

To sum up things : on the hype side there is deep learning, able to solve very complex problems using data when it is possible, and on the application side there are industrial problems where engineers need to have a global technical culture allowing them to pick up the right approach depending on the context... This is where AI is underexploited because engineers are not trained to that, and data scientists are very far from it... Education should integrate more data analysis with signal/image analysis and problem solving, and industry should have a non-dogmatic approach when dealing with "smart" applications (there are other tools than hammers - screwdrivers can perform well also)...