Explainable AI

An audience has to understand the reasoning or intentions of AI

One needs to understand the causes of an agents decision making

Explainable AI brings together various fields of research

• Social science

• AI

• Human-Computer interaction

To explain an event is to provide some information about its causal history. In an act of explaining, someone who is in possession of some information about the causal history of some event tries to convey it someone else

An explanation is a cognitive and social process

It answers the why-question

Correlation is a relationship between variables, where it is not completely clear which variable caused the other or if there is a third unknown variable which is the cause of them both

Causation exist when there is no other explanation for a circumstance than the potential reason behind it. It is a causal chain, a path of causes between a set of events. Correlation is necessary but not suficient

There are three levels of awareness

• perception of elements in current situation

• comprehension of the current situation

• projection of the future status

These lead to a decision

And then there is the performance of actions

Feature-based explanations –> AI tries to explain to which feature a certain unit belongs

Attribution-based explanations –> AI explains which regions of an image were important for the decision

Example-based explanations –> other prototypes or closest matches are shown

It is shown that XAI influences human decision making and that it improves the human decision making

People tend to trust XAI more than AI which does not explain there decisions

Increase trustworthines and informativeness

Ensure fairness and interactivity

Examine causality and transferablity

Domain experts and users of the model –> gain scientific knowledge and trust the model

Users affected by the models decision –> understand the reason and verify the dair decisions

Data scientists, developers and prodict owners –> improve product efficiency and gain insight in new functionalities or research