Bart de Boer (http://ai.vub.ac.be/members/bart) is in town and will give a talk on “Is unbiased learning possible?”, abstract below. Wednesday 10th October, 11am, Medical School, Doorway 4, Room G.16 (that’s approximately here: http://www.ed.ac.uk/maps/buildings/medical-school).

“Is unbiased learning possible?”

In the literature on Bayesian iterated learning, two different methods of going from a sample of observations to a production strategy are used: Maximum a-posteriori (MAP) learning and probability matching. This talk presents a mathematical argument that shows that both are impossible. The argument is based on a proof (by counterexample) that it is impossible to have an unbiased estimator of the mode (highest peak) of a continuous distribution. The impossibility of MAP learning and probability matching follows directly from this.

I will make an effort to present the mathematical argument in as accessible a way as possible. I will also discuss the implications of this finding, and argue that it makes the experimental search for (language specific?) learning biases more interesting.

• «Previous Post
• Next Post »

---