THE DEAD NEURON PROBLEM about bindsnet

[Hananel-Hazan]

Thanks for using BindsNET.
Is that an issue with BindsNET or do you have something to discuss that more suitable to the Discussions section?

[ZackWave]

感谢您使用 BindsNET。这是 BindsNET 的问题，还是您有什么要讨论的更适合讨论部分？

I used bindsnet to implement this network. After using postpre, only the part connected to input can send pulses, and the other parts will not work.

[Hananel-Hazan]

What network did you implement?

[ZackWave]

您实施了什么网络？

Network(
(inpt_layer): Input()
(M_layer): IzhikevichNodes()
(output_layer): IzhikevichNodes()
(inpt_layer_to_M_layer): Connection(
(source): Input()
(target): IzhikevichNodes()
)
(M_layer_to_M_layer): Connection(
(source): IzhikevichNodes()
(target): IzhikevichNodes()
)
(M_layer_to_output_layer): Connection(
(source): IzhikevichNodes()
(target): IzhikevichNodes()
)
)

[ZackWave]

您实施了什么网络？

input_data = poisson(datum=frame, time=time, device=device)
or
input_data = bernoulli(datum=frame, time=time, device=device)

[ZackWave]

What network did you implement?

This neural network can learn with Bernoulli but not with Poisson. It doesn't work even if made deeper. The learning rate is 1e-2.

[Hananel-Hazan]

I'm sorry, I dont understand what code do you use, and what is the issue you facing. From the little information that I have, you may have you may need to use higher weights values. But that is a guess.

[ZackWave]

对不起，我不明白你用什么代码，你面临的问题是什么。从我掌握的少量信息来看，您可能需要使用更高的权重值。但这只是猜测。

Thank you, the adjustment of wmax was effective.

[ZackWave]

对不起，我不明白你用什么代码，你面临的问题是什么。从我掌握的少量信息来看，您可能需要使用更高的权重值。但这只是猜测。

How should I set appropriate wmax, wmin, and norm values? And how can I achieve sparse connections? I have tried setting some weights to zero, but this method does not work due to postpre.

[Hananel-Hazan]

Some parameters like wmax, wmin and others needed to be tune by trial and error. About the sparse connection, currently PyTorch dont have full support for sparse matrices, therefore we cant use sparse connections in BindsNET. Hopefully the situation will improve in the future.

[ZackWave]

一些参数，如wmax，wmin和其他参数需要通过反复试验进行调整。关于稀疏连接，目前 PyTorch 没有完全支持稀疏矩阵，因此我们不能在 BindsNET 中使用稀疏连接。希望将来情况会有所改善。

How to implement STDP learning rule on inhibitory neurons.

[Hananel-Hazan]

The same as you implement the rule on excitation neurons just with negative weights

[ZackWave]

就像你只用负权重在激发神经元上实现规则一样

https://snntorch.readthedocs.io/en/latest/tutorials/tutorial_5.html

[Hananel-Hazan]

This is a nice paper but not biological realistic, since synapse polarity cant change in biology and BP can flip the sign. In BindsNET you can define the synapse in such a way that this event never happen. Please see the BindsNET documentation.