Replace two sequential conv layers with one

In theory, two circulant convolution layers(or more) should be replaceable by a single conv layer by assosiative property. Although circulant conv might introduce bad boundary artifacts, have anyone thought about using the assosiative property to reduce the size of neural nets?


Won’t the non-linearities prevent you from doing it?

Even without non-linearities, you could do it for evaluation, but the behavior for training would be different as you would have different degrees of freedom for your weights.