Humans are the drivers and benefactors of the progress in artificial intelligence and machine learning. On the other hand, artificial intelligence and machine learning are breaking ground for new kinds of user interfaces and user experiences. Closing this loop, that is, optimally combining the strengths of humans and machines, is one of the most interesting scientific questions at the moment. This course provides a look into state-of-the-art research at this intersection of probabilistic machine learning, cognitive models, and human-machine interaction.The course will be useful for students how are interested in:
- machine learning: A bottleneck for AI systems is that they don't understand humans.
- cognitive science: Understanding human behaviour is of interest as such.
- human-computer interaction: A bottleneck for intelligent interfaces is modeling human cognition and behavior and getting humans to understand the system.
All these fields share a common underlying challenge, for which the modelling of humans, machines, and their environment can provide the solution.
The course will include (1) high-level introductory lectures on the main topics, including computational rationality, reinforcement learning, probabilistic programming, and user modelling, (2) a project work for hands-on experience and a deeper look into a topic, and (3) select readings of research papers for a wider scope overview.
Prerequisites: Either (1) good understanding of probabilistic modelling/machine learning and programming experience (implementing machine learning algorithms) with interest in HCI/cognitive science/reinforcement learning, or (2) good understanding of HCI and some background in probabilistic modelling/machine learning with interest to learn more. The course is intended for advanced Master’s level or PhD level students. Courses providing relevant background include, for example, CS-E3210 - Machine Learning: Basic Principles, CS-E4820 - Machine Learning: Advanced Probabilistic Methods, CS-E5710 - Bayesian Data Analysis, CS-E4800 - Artificial Intelligence.
Passing the course (1 and 2 required for 3 ECTS, and in addition 3 for 5 ECTS):
(1) present (20 minute presentation) one paper, which will then be discussed (~25 min for discussion) and prepare a handout (1--3 pages including notes about the discussion after the presentation; returned to mycourses within the week of the presentation); possibly done in groups of 2-3 students depending on the number of students; required for passing the course
(2) learning diaries on papers (learning diaries should be returned to mycourses before the meeting where the paper will be discussed; learning diaries returned late will count only half point)
(3) project work, including short presentation of it (~10min or poster, to be decided based on how many students) and report (3-6 pages, returned to mycourses); project work can be done in pairs
For 3 ECTS: grade=min(5, number of learning diaries returned); 0 if no paper presentation done; all course work must be returned to mycourses by the end of period V, 18.5.2018, to count.
For 5 ECTS: grade=(2/5) * grade of project work + (3/5) * min(5, number of learning diaries returned); 0 if no paper presentation done (and 3 ECTS if no project work returned); all course work must be returned to mycourses by the end of period V, 18.5.2018, to count.
Schedule (all meetings on Mondays at 14.15-16.00 in T3, CS building, except on 9.4. the meeting is in T6):
19.2. Overview, practicalities; Introductory lecture: Interactive machine learning (Lecturer: Dr. Tomi Peltola)
26.2. Introductory lecture: Reinforcement learning (Lecturer: Prof. Ville Kyrki)
5.3. Seminar: Experimental design and Bayesian optimization for interaction
- Sections 1,2, and 5 of Eric Brochu, Vlad M. Cora, Nando de Freitas. A Tutorial on Bayesian Optimization of Expensive Cost Functions, with Application to Active User Modeling and Hierarchical Reinforcement Learning. https://arxiv.org/abs/1012.2599 [assigned: course staff]
Brochu et al. A Bayesian interactive optimization approach to procedural animation design. 2010, Proceedings of the 2010 ACM SIGGRAPH/Eurographics Symposium on Computer Animation. https://dl.acm.org/citation.cfm?id=1921443 [assigned: course staff]
12.3. Introductory lecture: Computational rationality I (Lecturer: Prof. Antti Oulasvirta)
19.3. Introductory lecture: Computational rationality II (Lecturer: Dr. Jussi Jokinen)
26.3. Introductory lecture: Probabilistic programming (Lecturer: Prof. Aki Vehtari, Dr. Henri Vuollekoski)
2.4. No meeting
9.4. Seminar: Computational rationality and reinforcement learning I. Note: Room T6.
- Gershman, Horvitz, Tenenbaum. Computational rationality: A converging paradigm for intelligence in brains, minds, and machines, Science, Vol. 349, Issue 6245, pp. 273-278, 2015. http://science.sciencemag.org/content/349/6245/273 [assigned: Pradeep]
Lewis, Howes, Singh, Computational Rationality: Linking Mechanism and Behavior Through Bounded Utility Maximization, Topics in Cognitive Science 6 (2014) 279–311. http://onlinelibrary.wiley.com/doi/10.1111/tops.12086/full [assigned: Buse, Karina]
16.4. Seminar: Computational rationality and reinforcement learning II
- Howes et al., Interaction as an emergent property of a partially observable Markov decision process. Chapter 10 in Computation Interaction, edited by Oulasvirta, Kristensson, Bi, and Howes, Oxford University Press, 2018. [assigned: Thomas, Ville]
Rabinowitz et al. Machine Theory of Mind, https://arxiv.org/abs/1802.07740 [assigned: Janin] (
Kangasrääsiö et al. Inferring Cognitive Models from Data using Approximate Bayesian Computation, CHI2017,https://dl.acm.org/citation.cfm?doid=3025453.3025576)
23.4. Seminar: Probabilistic programming
- Cusumano-Towner et al., Probabilistic programs for inferring the goals of autonomous agents, https://arxiv.org/abs/1704.04977 [assigned: course staff]
Owain Evans, Andreas Stuhlmueller, Noah Goodman. Learning the Preferences of Ignorant, Inconsistent Agents, AAAI2016. https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/view/12476 [assigned: Mauro, Tinka]
30.4. Seminar: Compositionality, causality, and learning to learn I
- Lake, Salakhutdinov, Tenenbaum, Human-level concept learning through probabilistic program induction. 2015, Science. http://science.sciencemag.org/content/350/6266/1332 [assigned: Kunal, Stig-Arne]
Santoro et al, Meta-Learning with Memory-Augmented Neural Networks, ICML2016, http://proceedings.mlr.press/v48/santoro16.pdf [assigned:] - there will probably not be a presentation of this paper
7.5. Seminar: Compositionality, causality, and learning to learn II [note: these papers will be assigned to 2 groups of students, who decide how to split the presentation between them (that is, split the first paper in a suitable way and second group takes additionally the second paper); assigned: Fabio]
- Lake et al. Building Machines That Learn and Think Like People, Behavioral and Brain Sciences, 40, E253, 2017. https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/building-machines-that-learn-and-think-like-people/A9535B1D745A0377E16C590E14B94993
Botvinick et al. Building Machines that Learn and Think for Themselves: Commentary on Lake et al., Behavioral and Brain Sciences, 2017, https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/building-machines-that-learn-and-think-for-themselves/E28DBFEC380D4189FB7754B50066A96F
Final meeting: project presentations, summary, feedback on course No meeting!
Course organizers: Tomi Peltola, Mert Çelikok, Samuel Kaski; Probabilistic Machine Learning research group
Please, use the course forum for general questions and issues regarding the course. The course staff can also be contacted by e-mail if necessary.