- IHDM - https://arxiv.org/abs/2206.13397
- FlexConv - https://arxiv.org/abs/2110.08059
From FelxConv:
Intuitively, multiplication with the mask blurs in the frequency domain, as it is equivalent to convolution with the Fourier transform of the mask
Feel like this would pair nicely with IHDM since that's basically exactly the process that's being modeled
I think the CKConv repo has FlexConv? https://github.com/dwromero/ckconv
In FlexConv, they regularize the kernels highest frequency to mitigate aliasing, but I feel like they could take this a step further by applying a temperature/schedule to this regularization, starting at a frequency lower than the nyquist frequency to produce a low resolution kernel, than increasing the highest allowed frequency up to nyquist as learning progresses, so the learning momentum is encouraged to identify a direction that sharpens the kernel resolution as learning progresses