Make small changes to the inputs and check its effects on the output. If the small change has a high impact, we can deduce (Not guaranteed but likely) that these inputs are very relevant to the output predictions.

This method does not utilize the architecture of the neural network for its explanations and is therefore "limited".

A different approach, that does utilize the architecture of the neural network for its explanation is LRP