Learning Multiple Categories on Deep Convolution Networks.

Deep convolution networks have proved very successful with big datasets suchas the 1000-classes ImageNet. Results show that the error rate increases slowlyas the size of the dataset increases. Experiments presented here may explainwhy these networks are very effective in solving big recognition problems. Ifthe big task is made up of multiple smaller tasks, then the results show theability of deep convolution networks to decompose the complex task into anumber of smaller tasks and to learn them simultaneously. The results show thatthe performance of solving the big task on a single network is very close tothe average performance of solving each of the smaller tasks on a separatenetwork. Experiments also show the advantage of using task specific or categorylabels in combination with class labels.