ML and Industry

There is widespread belief that many Machine Learning methods are only efficient for "cats and dogs" images but not for other industrial "complex" data. It can be partially true but there are other important facts to pay attention to:

1- If we speak mathematically or geometrically, patterns that exist in these data can be sometimes more complex than other patterns in many industrial data. The truth is that our powerful human vision system makes sometimes their interpretation more intuitive for us. This helps, in a large part, to design powerful algorithms, models and data structures for better computer-based interpretation/comprehension.

2- AI/ML researchers spent decades handling, most of the time, computer vision problems (for many reasons, it is another topic). This doesn't make them simple today. Moreover, many CV problems are not yet resolved...

3- We have to get out from this "data type"-based categorization of AI/ML (text, image, video, audio, ...). It is rather a question of data representations, problem modelling, deep knowledge of the implementation platform (hard/soft pos and const) and most of all the coherence between these components.

4- We cannot directly assume that ML models designed basically for data that are interpretable by natural human being sensors/systems (vision, language, audio, etc.) to be obviously efficient for other industrial data with different physical dynamics/nature ... That is why THERE IS NO OFF-THE-SHELF AI.

5- More space and time must be given sometimes to researchers to rethink the whole ML pipeline from scratch for SOME industrial problems. It will lead not only to find the right solution or the right scientific answer but also to make meaningful advances in AI/ML the same way computer vision applications did for AI decades ago.

6- For all previously mentioned point and speaking about AI, that is why R&D is essential in any industrial sector.

Sonia Yousfi