Keynote@FOIS'24 - Where to locate the explainability of explainable machine learning? By Mieke Boon

New title: Exploring AI/MLR Epistemology

Epistemology 16th-17th century
- Rationalism - Rene Descartes
- Empiricism - Francis Bacon

Thinking about Empiricism (which resembles ML processes), there were criticism:

Hume (1748) An enquiry concerning Human Understanding
Threre are fundamental problems:
- induction: the principle of induction is logically invalid;
- causality: causal connection, e.g., the power, cannot be observed.

Practical example of (lack of) causality:

Further developments in Empiricism:

Problems with ML empiricism:
Problem 1 - If you look for patterns in the data, you will find them (even if there is no causation)
Problem 2 - Need for explanations

How does (Logical) empriicism solve the problem of explanation? - without causes or mechanism (anti-metaphysics)

The way it is looked at today in ML, explanations are similar to what we do in the lab: this variable has a lot of effect in the result, while this one does not. Problem 3 - It denies the epistemic and pragmatic value of (causal)-mechanistics explanation: explanations of regularities.
Philosophers of science criticize empiricist epistemology and aim at solutions
- Alternative epistemology should answer: what is a real law? How do we know that a mathematical structure or statically relevant correlation found in the data is a real law? Response: iff there is a mechanism that explains the law.
- the mechanism thus makes the law intelligibel - i.e., it explains the law
- human reasoning in science: rather than identifying laws, researchers explain by constructing a model of the underlying mechanism.

Kant's epistemology (18th centruy) : concepts + power of jusdgement Kant reconciled and transcended the rationalist and empiricist epistemologies. by providing an alterrnative to the traditional questions of: what is the baseis of true knowelge? how can we be certain?
Kant's questions: How is it possible that we have knoweldge of the world? What are the conditions for the possibility of having knowelvge anyway?

Kant claims:
- Man himself creates all his respresnetations and concepts. - Concepts as conditions for the posbiiility of having anc crating knowlefge about reality. withouth these concepts, we would not be albe to make any staetemnt about reality on the bases of mere observations.
- Perceptions without concepts are empty; concepts without perceptions are blind.
- Kant considers the crucial and intellegcuatl role of human in creating concepts.

Kantian Epistemology => Conceptual Modeling

She presents a High school level in which conceptual modeling precedes mathematical modeling:

Kant:
Concepts (verstand) + power of judgement (urteilskraft) - meaning, values appreciation / emotion.

How do we make sure the model is a good representation of the world?
- Picture of the pope is a good model of the Pope.
- ... But not of Trump
However, Trump is also seating in a chair in the same position as the pope! So, in the world, a model can be similar to the world in many different ways.

In summary...