Closed pseudo-rnd-thoughts closed 3 weeks ago
Hi, thanks for your question! Actually, we used the definition in The Dormant Neuron Phenomenon in Deep Reinforcement Learning, which is the paper that "designed" dormant ratio. I agree that there might be a larger percentage of neurons are dormant than the paper suggests using this definition. Therefore, in the follow-up work ACE, we changed the definition to "gradient", which might solve this issue.
Thanks, I hadn't spotted that. Good luck with the paper
Thanks for open sourcing your work
I was looking through your code, in particular, the
cal_dormant_ratio
function. You compute the absolute value of the forward pass of observations, why? As you are using ReLU with your linear layers then for negative values, these have an activation of 0 not positive given absWas this intentional?
If not, this would implement that a larger percentage of your neurons are dormant than the paper suggests