Closed wusaifei closed 4 years ago
Hello, SNNs are super sensitive to the network configuration and parameter values such as number of layers, number of features, receptive sizes, and learning rates to mention a few. Thus, you need to tune everything for each new dataset in order to get the best outcomes.
@miladmozafari Thank you very much for your reply. How about changing the weight update to t-stdp(Triphasic spike-timing-dependent plasticity)? I don't think the author has considered the influence of time difference? Try to change the weight update formula. It is found that it cannot be changed to t-stdp rule because there is no time parameter in it. Does the author have any good ideas?
I think implementing t-stdp with this framework is not straightforward and needs some hacks. We represent time with the batch dimension, meaning that each tensor along the batch dimension stores data of that particular time-step.
@miladmozafari Does the author consider adding speech recognition?
It's not planned for SpykeTorch, but see https://github.com/romainzimmer/s2net
@tmasquelier Thank you for your reply。
@wusaifei Were you ever able to find parameters that improved accuracy on CIFAR-10? I have been trying for weeks with no luck.
Hello, I changed the training data into cifar10 and then cifar10 into gray-scale image, but the accuracy is not high, only 36%. The general CNN network has an accuracy rate of more than 66%. Why?