Text this: Visual tasks beyond categorization for training convolutional neural networks