Bad Convergence in Machine Learning

Machine learning has recently been revamped and redesigned with well abstractly defined goals. Tagged to these goals is the vision of bringing ML to everyone, with any background. While this has so much effect in the transformation of research in different areas, and transcendence of the understanding of understanding of humans by computing machines, it introduces a lot of bad optima convergence, it that sounds right.

Machine learning is simply the use of computing capabilities of machines to run models that we think approximate the human thinking capacity in solving problems. Since these models are generic, a lot of disciplines (from engineering to humanities) can leverage these computing capabilities to speed up computation over their domain specific models. However, since the design and implementation of these models rely on rigorous mathematics and computer science principles (from computing framework, algorithm design and testing), and not every domain/discipline covers these, the following results:

1. Computer scientists, together with software engineers and mathematicians collaborate to design tools that can be used to perform machine learning without an in-depth understanding of the computing framework and the algorithms. Often a language that is easy to learn is used (such as Python, and other untyped syntax languages). These framework include Goolge's TensorFlow, Facebook's PyTorch, nd others such as Keras, DeepLearning4j, etc.
2. Researchers from various backgrounds can then learn the high level APIs and frameworks that are produced in (1) and be able to perform machine learning without having to stress about the math and the underlying (or back-end) code.
3. With research organisations awarding publications, and often quantity over quality (in most cases), these researchers end up using the high level abstractions to derive research publications, and as a result, this trend converges to bad optima where these individuals begin to believe and be treated as machine learning specialists/experts. It can only be great for the individual.
4. As a result of this convergence, the core science remains with the few that understands it, and the application expands and spans larger spaces. The expansion is a great response from the deep research done by the few that understands it. However, from (3), these individuals begin to transform into "non-expert experts", and possible land in leadership positions in large ML research units/labs. This is what may not be great for the few researchers that understands ML very well.

From this point on, anyone can project a conclusion of this post. I live it to the reader to infer on from their own intuition. The point I wanted to display here is that:

Data is doing us a great job in understanding the world around us without having to manually interact with it, and it would only benefit us to take relevant and rigorous courses in machine learning maths and computer science, so as to enhance and distribute ourselves towards the real and impactful research that transcends the field.

We can never understand the brain's thought flow/patterns by mining abstractions over them. We have to go deeper, granular, and build more lower abstraction which we should also break down and mine their constituents.

So since you know where I stand, please hit me up if you'd like to collaborate.