Reddit - MachineLearning - [D] ReLU intuition with the dot product?

I came to thinking about another benefit for ReLU, which deals with the "thresholding" behavior of the ReLU. My interpretation has to do with the dot product in convolutional neural networks. As everyone knows, when the kernel is slid over its input window, it computes the dot product as with vectors. The process of convolution can be interpreted as looking for a specific pattern throughout an image. Now, the only thing that affects the dot product's sign is the angle between the two vectors--in this case, the kernel and its window.