In this course, students will gain knowledge on (Bayesian) inference to the best explanation, both from a conceptual, philosophical, and computational perspective. Students will learn about (and become aware of) some of the fundamental problems that autonomous cognitive agents (such as humans, robots, or software agents) encounter when they need to infer explanations for the phenomena they observe, such as ‘what is a good explanation', ‘how to decide what is relevant', and ‘how can it be done tractably'. In particular, students will learn to combine, integrate, and contrast philosophical approaches to inference to the best explanation with computational approaches in Bayesian networks.
|
|
An embodied and embedded autonomous agent - such as an exploration device dropped on Mars - needs to make sense of what it encounters through its sensors, for example, to decide what to do next. That 'making sense' is very broad, ranging from deciding upon what is relevant, to generating and selecting candidate hypotheses, to deciding upon which is the best explanation. Importantly, it needs to make such decision with limited resources. We want this agent also to explain and motivate its decisions to its users. In this course we try to 'make sense of making sense' and 'explain explaining': we discuss the fundamental problems that such an agent needs to solve. In particular, we will focus on Bayesian inference to the best explanation.
The course consists of four blocks and a final specialization project. In the first block, we focus on the philosophical perspective of inference to the best explanation. For example, we will look at different proposals for what constitutes 'best' in 'inference to the best explanation' and we will look at Fodor's and Dennett's conception of the Frame Problem.
In the next block we go into depth with respect to the theory and foundations of uncertainty and stochastic computations, including its computational complexity. In the third block we introduce the MAP problem in Bayesian networks, i.e., the problem of finding joint value assignments with maximum posterior probability given evidence in the network. In the final block we will discuss several computational models, based on recent literature, that describe 'relevance', 'informativeness', and 'counterfactual reasoning' within a Bayesian framework.
The course will be examined with a written open book exam, a specialization project, and the project proposal. The specialization project builds on one of the topics of the course (and the instructor's expertise and research program), and may depend on the student's particular background and research interests. Typically this results in a literature review that deepens the student's knowledge on one of the topics of the course, or a small research project, e.g., robot experimentation, algorithms design, complexity analysis, or conceptual/philosophical analysis are also possible. The specialization project can be done individually or in a small team.
|
|
|
|
A completed (academic) BSc degree in Artificial Intelligence or related field, such as Cognitive Science, Philosophy, or Computer Science. The course assumes that students have sufficient background in both more computational and more philosophical aspects of AI. The course expects some basic knowledge of Bayesian networks and computational complexity theory; a self-study guide is available for students that lack this background. |
|
- Written open book exam (60% of the final grade)
- Project proposal (10% of the final grade)
- Specialization project (30% of the final grade, includes project report)
All parts need to be marked at least 5.0 (with the weighted average > 5.5) in order to pass the course.
|
|
This is a core course in the AI:Cognitive Computing specialization and a specialization elective in the AI:Intelligent Technology specialization. The course is open for master students in philosophy, computer science, cognitive neuroscience, or a similar programme.
|
|