I was wondering how specific symmetries or constraints such as property-invariance transformation are imposed on any (deep) neural net when they are trained.

I'll appreciate it if anyone can aware me of the relevant researches in this direction.

  • $\begingroup$ Do you have specific "properties" in mind? Convolution NNs are famously translation equivariant. Achieving other symmetries is harder: arxiv.org/pdf/1909.12057.pdf $\endgroup$
    – bogovicj
    Jun 29, 2020 at 19:04
  • $\begingroup$ Thanks for the paper! Shift-invariance was the first example that occurs to me. In fact, in terms of object detection and in particularly the convolutional nets, I believe any Affine transformation should in principle keep the result intact, although there might be no guarantee on better performance. $\endgroup$
    – arash
    Jun 30, 2020 at 21:00


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.