About generalization, abstraction and analogies
A presentation of essential cognition concepts inspired by the book Deep Learning with Python, second edition by François Chollet.
Introduction
The pursuit of better generalization is probably the underlining target of most machine learning efforts, both for research and industry as well.
The basic idea is to build a model that is capable of handling unseen data, the more different this data is from the set of samples that the model was trained on, the more difficult this task becomes. However, we are always working in a rather small area of generalization that is representative of the task we are trying to solve, we are trying to build good but extremely limited models.
Specializations levels
By looking at the variety of problems and data in the machine learning scene we can try to create a hierarchy of specialization:
- The first level is the closest one to the concrete problem we are tackling, let’s take an object detection task, we want a model that is able to localize and classify the type of objects we are interested in (people, cars, animals, etc.). If some problems are similar enough we can of course use transfer knowledge and reuse some of the layers between them but that is also limited by the similarity…
