Encoding Reusable Perceptual Features Enables Learning Future Categories from Few Examples
Michael Fink, Kfir Levi · 2004
A perceptual system coping with a dynamic environment must be able to learn to detect new object categories from a few examples. However, learning from a small sample is restricted by the hindering effects of model overfitting. We present an algorithm aimed at circumventing the effects of overfitting by utilizing a set of reusable features, learned from several previously trained categories. We show that when applying a feature reuse strategy the algorithm learns complex real world objects from few examples.