Putting Bayes to Sleep

In online multitask learning a single algorithm faces a collection of interleaved tasks. By addressing these tasks jointly rather than in isolation, the algorithm can discover and exploit similarities between tasks and hence learn faster and perform each task better. We present a new method for multitask learning, built atop the "specialist" framework. We obtain a new intriguing efficient update that achieves a significantly better bound. Our method has linear efficiency and hence scales to very large data sets.