What does it mean when we say an algorithm/metric is agnostic




I have all kinds of machine learning terms that co-occur with the word "agnostic", including model-agnostic learning, model-agnostic metric. From the dictionary, it explains the word "agnostic" in the following way

a person who holds the view that any ultimate reality (such as God) is unknown and probably unknowable.

which does not make those terms more understandable.

In some contexts, I find that "agnostic" refer to "generic" or "free of". For example, in the paper I am reading now, the authors define a threshold-agnostic metric, where they use score rather than hard 0/1 assignment for the task.

However, I am wondering if there is formal definition for the word "agnostic" in the machine learning community.


Posted 2019-10-09T05:31:57.817

Reputation: 121



I know of the model-agnostic term, and a close meaning would be model-independent. Basically, when you are studying a machine learning problem, the underlying structure in the data may or may not be described by one type of model or the other. The model-agnostic approach consists in using machine learning models to study the underlying structure without assuming that it can be accurately described by the model because of its nature. This avoids introducing a potential bias in the interpretation.

A model-agnostic approach pretty much requires that several different techniques are used for the same task.

Romain Reboulleau

Posted 2019-10-09T05:31:57.817

Reputation: 1 207

Thank you for your answer. Could you provide a pointer to a concrete example so that I could fully understand "model-agnostic"? – Mr.Robot – 2019-10-09T06:05:44.623

If you have one specific model in mind, say a decision tree, you are sticked to the interpretation that data can be classified using a restricted set of rules (in that case, a combination of binary tests). This may be right, and you can perfectly do that if you have good reasons to believe that it works. On the opposite, model-agnosticism means trying to find an explanation of how the data is ruled, without making such assumptions that rely on model choice. Hope that helps. – Romain Reboulleau – 2019-10-10T03:10:42.680


"Agnostic" means that nothing is known or can be known of the existence.

In data science, the term "Agnostic" is used when there's self-modelling/self-learning technique is involved.

For e.g.

A robot modeled itself without prior knowledge of physics or its shape and used the self-model to perform tasks and detect self-damage [1]. It is task agnostic self modeling.

It is reminiscent of meta learning.

Nischay Namdev

Posted 2019-10-09T05:31:57.817

Reputation: 141