From hard programming to hard training: are we really evolving?

One of the great benefits of machine learning (ML) is related to the fact that we can create models as or more complex than those developed in the form of algorithms by experienced programmers, often through hours and hours of programming. However, deep learning (DL), which can be considered the state of the art of ML and its most successful case, can lead to long adjustment and training processes, due to the large number of layers and parameters of artificial neural networks, also involving hours and hours of work. In addition, because automatically created models do not accompany interpretation or semantics, for most of the cases, what we have in practice is mostly a black box of high complexity and effort to address if anything behaves out of the expected.

And if this is the reality for many complex problems, I ask: are we really evolving?

My answer is that we should change, as soon as possible, and focus on evolving, especially in relation to efforts for interpretation. In fact, I see no logic in changing the intelligence of programming, much closer to interpretation, by a manual effort of representation that is not accompanied by better results.

And how to evolve in interpretation?

Simple, dedicating at least the same time, and the current investment, in the production of models automatically for production of models interpretation. That is, I believe that organizations that invest heavily in AI should think about the future, reserving significant space for reverse-engineering the generated models, and probably the paradox is that this is a job for data science teams with advanced programming knowledge.

In addition, I believe that programming efforts, which can lead to strong AI, should not be replaced, but only complemented, although in practice, all the ease and evolution of agile methods, in line with deep learning, is to completely replace programming processes.

And, of course, with the advances in ML technology, mainly of generalization and interpretation of the models, these investments are expected to be smaller, if we really find the path of wisdom and evolution.

But, as long as we do not reach this level of evolution, we need to create spaces to actually be worth replacing programming with training.

By Rogerio Figurelli at 06/07/2020

Leave a Reply

Your email address will not be published. Required fields are marked *