Opening the algorithm’s black box and understand its ouputs

摘要

In this article, we will define the principles of loyalty and watchfulness in algorithm’s usage. We will then explore some of the interpretability techniques, which let us explain the decisions of an algorithm and share them with transparency. We will focus on ways to express indecision in order to increase the trust of the users in those systems. The second part is dedicated to a collection of statistical biases, which can be introduced both in the conception and the usage of those algorithmic processes. We will define a methodology to control the possible outputs (decisions) of an algorithm and keep them from conflicting with the interests of the algorithm designer and users.