Machine learning (ML) mainly answers such questions as “What?” / “Who?” / “How much?” as well as "What is depicted?" t and so on

Why interpretations are necessary

submited by

Style Pass

2021-06-17 18:00:10

Machine learning (ML) mainly answers such questions as “What?” / “Who?” / “How much?” as well as "What is depicted?" t and so on. The most natural human question that follows is “Why?”. In addition to the black box answer (whether it be boosting or a neural network), we would like to receive the argumentation of this answer. Below is an overview of the interpretation problem.

There is no precise definition of the interpretation of the model for obvious reasons. However, we can all agree that the following problems lead to the necessity for interpretation:

One quality (model performance) metric (and even a set of metrics) does not describe the behavior of the model, but only the quality itself on a specific sample

The use of machine learning in “critical areas” (medicine, forensics, law, finance, politics, transport) gave rise to requirements for the security of the model, justification of confidence in it, as well as various procedures and documents regulating the use of ML-models

A natural requirement for an AI that is capable of answering the “What” question to justify the "Why" one. For example, children begin to ask this question even before they receive any in-depth knowledge. This is a natural stage in cognition: if you understand the logic of the process, then its effects and consequences become more obvious.