Learning with Few Examples using a Constrained Gaussian Prior on Randomized Trees.
Erik Rodner, Joachim Denzler · 2008
Machine learning with few training examples al-ways leads to over-fitting problems, whereas human individuals are often able to recognize difficult ob-ject categories from only one single view. It is a common belief, that this is mostly established by transferring knowledge from related classes. There-fore, we introduce a new hybrid classifier for learn-ing with very few examples by exploiting interclass relationships. The approach consists of a random-ized decision trees structure which is significantly enhanced using maximum a posteriori (MAP) esti-mation. For this reason, a constrained Gaussian is intro-duced as a new parametric family of prior distri-butions for multinomial distributions to represent shared knowledge of related categories. We show that the resulting MAP estimation leads to a simple recursive estimation technique, which is applicable beyond our hybrid classifier. Experimental evaluation on two public datasets (including the very demanding Mammals database) shows the benefits of our approach compared to the base randomized trees classifier. 1