CSNN loss function for image input to image output [2D -> 2D]

[Radu2k]

Hello,

I am working on an implementation of a CSNN network using rate encoding but the output is an image, so the output is
[num_steps, batch_size, out_size1, out_size2].

I was looking around for a loss function but the only one that worked for the first two iterations was Cosine Similarity. As problems appear when backpropagating due to the fact that the output is unsqueezed to compress my rate encoding I was wandering if any of the SF have this sort of functionality?

Thanks!

[jeshraghian]

Hi Radu, We haven't natively implemented any losses yet into SF that account for problems outside of classification, though it's definitely an important feature. I suppose the loss would need to manually constructed. Eg, sum up the spikes in the output, divide it by the number of time steps to normalise everything to 1, and then apply cosine similarity or mean square error to compare it to the original image. There is some recent work on SNN autoencoders, so I'll look into those to see how SF can be expanded to handle losses acting on multidimensional outputs.

…

On Wed, 16 Nov 2022, 7:54 am Radu Salavat, ***@***.***> wrote: Hello, I am working on an implementation of a SCNN network using rate encoding but the output is an image, so the output is [num_steps, batch_size, inp_size1, inp_size2]. I was looking around for a loss function but the only one that worked for the first two iterations was Cosine Similarity. As problems appear when backpropagating due to the fact that the output is unsqueezed to compress my rate encoding I was wandering if any of the SF have this sort of functionality? Thanks! — Reply to this email directly, view it on GitHub <#145>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJTFT4R6GIX2UIEXEYYUUWLWIT7T7ANCNFSM6AAAAAASCMLLRM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Radu2k]

Hi Jason,

That sounds good to me. Looking forward to seeing this implemented. If there is any way to help implementing this please let me know.