Why does a Convnet not work for sequence modeling?

[electro]

Based on the above slide, I understand that inputs/outputs may be of varying size, but assuming we ignore this issue, I don’t understand the second bullet point…

Could we not make a convolutional network with input dimensions of (m x Tx_max x 10000 x 1)? If so, what prevents this Convnet from sharing features across different positions, and in that case why do normal Convnets succeed in sharing features across positions of an image?

[TMosh]

The standard NN uses the linear combination of the examples and features in each layer - so the order of the examples does not matter.

For a sequence model, the order of the examples is critical.