Goals of the meeting
- Build probabilistic models of humans, infers them with likelihood-free inference, and use the models to revolutionize HCI.
The basic idea of the human-in-the-loop Approximate Bayesian Computation (ABC) is to interactively query the user about the simulation/calculation results and Bayesian optimization. In this way, rather than using some machine learning algorithm, a human covers the role of the classier for the ABC procedure. There are at least two relevant applications in which the human classier can improve the efficiency (in terms of computational time and/or total number of draws needed for covering the posterior distribution) of the ABC algorithm and the achievement of the true posterior distribution as well:
- Elicit prior knowledge.
Address the problem related to the selection of the discrepancy measure/summary statistics.
We would ultimately want to learn a joint model of the user and the task, such that we can then use the joint model to design interaction with the user. This can be done either by learning jointly, or by learning separately a user model. The classier ABC would map to this problem by asking the user feedback about the plausibility of the outputs of the user model.
Discussion items / In meeting
Action items for the next meeting
Next Meeting in A346 Otaniemi + skype on February 7th 9:45 am - 11 am
According to what was introduced in this kick-off, now is time to think at feasible examples to introduce and develop the project. A suitable case study will include the user model, the world model and the interactive AI. When moving through these three steps, we could consider incremental levels of difficulty.
- According to what was introduced in this kick-off, now is time to think at feasible examples to introduce and develop the project. A suitable case study will include the user model, the world model and the interactive AI. When moving through these three steps, we could consider incremental levels of difficulty.
- Umberto: Introduction on Astrostatistics and presentation of a possible scenario
- Tomi: Elicitation of non-linearity
- Pedram: Intent modelling / retrieval
- Antti: Model selection for linear models
- Sami: Personalized medicine
- Zhirong: User model for visualization