search

KEDRI-AUT / snn-encoder-tools

Data encoders
MIT License
31 stars 14 forks source link

Questions about BSA encoding #4

Open lovingpolar opened 1 year ago

lovingpolar commented 1 year ago 
Hello, I read your paper ’Selection and Optimization of Temporal Spike Encoding Methods for Spiking Neural Networks‘ to learn about BSA encoding. 
I'm trying to do a grid search with the whole dataset to find the optimal filter parameters, but I find it strange that the lower my cutoff frequency is, the higher my SNR is.This prevents me from getting an optimal parameter.How can this be solved?
petrobalint commented 1 year ago

Hello, It would be helpful to see an example waveform and/or frequency spectrum of a sample of your data. Firstly, keep in mind that the cutoff of the filter should match the upper bound of the spectrum you want to keep in your data. Secondly, it could be the case that SNR is not the best for your case. Have you tried looking at the RMSE? You could easily change the code to calculate RMSE during the search instead of the SNR. Another interesting observation is that the spiking rate is around (or usually a bit lower than) 0.5 when the encoding is considered optimal. Best, Balint

gdgy @.***> ezt írta (időpont: 2022. nov. 18., Pén 13:20):

Hello, I read your paper ’Selection and Optimization of Temporal Spike Encoding Methods for Spiking Neural Networks‘ to learn about BSA encoding.

I'm trying to do a grid search with the whole dataset to find the optimal filter parameters, but I find it strange that the lower my cutoff frequency is, the higher my SNR is.This prevents me from getting an optimal parameter.How can this be solved?

— Reply to this email directly, view it on GitHub https://github.com/KEDRI-AUT/snn-encoder-tools/issues/4, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJOYGX4CU4KKPW4IRX3NO53WI5YBZANCNFSM6AAAAAASEN3MG4 . You are receiving this because you are subscribed to this thread.Message ID: @.***>

lovingpolar commented 1 year ago

Thank you very much for your reply. Regarding the first question, I use 5 types of data in the dataset, if the cutoff frequency needs to match the upper bound of the spectrum, does that mean that I need to encode each type of dataset separately ？(I currently intend to unify the filter parameters for the entire dataset) Regarding the second question, I have also tried to use RSME as a encoding criterion before, probably because I did not pay attention to the filter cutoff frequency matching the upper spectral bound of the signal, so the result is similar to using SNR to optimize.I also found that the smaller the cutoff frequency, the larger the SNR, the less detail the waveform. Finally, I will also try to calculate the rate of spike signal

lovingpolar commented 1 year ago

image image This is the waveform and spectrum of a sample of two types of data