Text this: Speeding up convolutional neural networks with low rank expansions