Request for Guidance: Applying Liquid Time-Constant Concepts to Spiking Neural Networks

[ypl315654]

Dear Dr. Hasani,
I was reading your paper on Liquid Time-Constant Networks and was particularly intrigued by the method mentioned where inputs are multiplied by A-v to simulate the suppression of new input signals as the cell membrane potential increases. I attempted to apply this concept to the Leaky Integrate-and-Fire (LIF) model within a spiking neural network framework. The neural dynamics formula for the LIF model I employed is attached in the file.
In experiments conducted on the cifar10dvs dataset, I observed that incorporating this suppression effect resulted in a decrease in final test accuracy compared to the original neuron model. Could you advise if there are any specific considerations I should be aware of when adapting your ideas to spiking neural networks? Or, is it possible that such suppression due to increased potential is not applicable in image processing tasks? Thank you for your time and assistance.

Yours sincerely,
A student from the Shanghai University.