Classifier adaptation at prediction time

Classifiers for object categorization are usually evalu-ated by their accuracy on a set of i.i.d. test examples. This provides us with an estimate of the expected error when ap-plying the classifiers to a single new image. In real appli-cation, however, classifiers are rarely only used for a sin-gle image and then discarded. Instead, they are applied se-quentially to many images, and these are typically not i.i.d. samples from a fixed data distribution, but they carry de-pendencies and their class distribution varies over time. In this work, we argue that the phenomenon of correlated data at prediction time is not a nuisance, but a blessing in disguise. We describe a probabilistic method for adapting classifiers at prediction time without having to retrain them. We also introduce a framework for creating realistically dis-tributed image sequences, which offers a way to benchmark classifier adaptation methods, such as the one we propose. Experiments on the ILSVRC2010 and ILSVRC2012 datasets show that adapting object classification systems at predic-tion time can significantly reduce their error rate, even with no additional human feedback. 1