a question about your work in conversion of ANN to SNN

[Xingfush]

Hello, I am trying to reproduce the result in paper theory and tools for the conversion of analog to spiking convolutional neural networks at Matlab platform and got a disturbing questions: The dynamics of output neurons is severely influenced by large biases. As simulation goes, the output result converge to neuron with largest biases. I have tried 'voltage clamp' method raised in your paper, but didn't got expected effect (little improvement). Thank you very much for your work in this paper. I have been confused by first problem for long time. Looking forward to your reply!

[Xingfush]

I have tried all tricks raised in your previous two papers, such as spiking max pooling, spiking softmax, analog input and normalization scale. They all achieved good conversion loss reduction. But once I add biases in network, the accuracy of SNN will be very bad. The classification result will converge to the output neuron with biggest bias. Also, I find the biases in fully-connected layer rather than convolutional layer cause this problem, as you talk in your paper.

[rbodo]

Hi,

Yes, large biases are a problem in SNNs. There is no simple way around it. We explored a number of options:

• Train the ANN without biases.
• Clamp the membrane potentials to zero for a number of time steps.
• Initialize the membrane potentials with a value that is proportional to the bias, multiplied with a minus sign.
• Fix the biases to zero at the beginning of the simulation, and slowly increase them to their true value. It may be that only the first option removes the problem completely.

Best,

Bodo

On Sat, Jul 14, 2018 at 9:07 PM Xingfu notifications@github.com wrote:

Hello, I am trying to reproduce the result in paper theory and tools for the conversion of analog to spiking convolutional neural networks at Matlab platform and got a disturbing questions: The dynamics of output neurons is severely influenced by large biases. As simulation goes, the output result converge to neuron with largest biases. I have tried 'voltage clamp' method raised in your paper, but didn't got expected effect (little improvement). Thank you very much for your work in this paper. I have been confused by first problem for long time. Looking forward to your reply!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/NeuromorphicProcessorProject/snn_toolbox/issues/15, or mute the thread https://github.com/notifications/unsubscribe-auth/AO9TCaZW5VsThJ4FtsfyFtx1YJq8JKupks5uGrH2gaJpZM4VQGDV .

-- Bodo Rückauer Kronwiesenstraße 68 8051 Zürich Switzerland +41 (0)78 882 41 61

[Xingfush]

Thank you very much, I'll try these methods you provide and make more exploration.