Disclaimer: Dieser Thread wurde aus dem alten Forum importiert. Daher werden eventuell nicht alle Formatierungen richtig angezeigt. Der ursprüngliche Thread beginnt im zweiten Post dieses Threads.
Finding bias and random weight
I have implemented the neural network from the current assignment and it is working so far. Now I was adjusting the values for bias and the random weight and was wondering how I can find “an appropriate constant value for bias”.
Therefore I tried different values and sometimes every testcase is correct but sometimes they aren’t. I have found a constant value that fits to the network but this can’t be the way it should work. How can I calculate an appropriate bias from the input parameters? At the moment I don’t use them. If this is necessary, I would ask the same question for the random weight, because I’m currently using Math.random(). :rolleyes:
That seems appropriate, given that they’re supposed to be random 
I’d defer to Marcel’s judgement, but as far as I know, fine-tuning the parameters of a neural network is more like an art than a science - i.e. not something we can do algorithmically, but rather something that is done using a lot of domain-specific heuristics…
That is exactly (and unfortunately) the way it works. There are a few general rules of thumb, but nothing which will guarantee you in every situation the best result possible. I think understanding this is one of the key moments when working with Neural Networks. Everyone talks about how easily they outperform even humans, but barely one talks about the extreme amount of domain-knowledge needed to create a NN, which works that good.
So yep for you probably „guessing good distributions (or googling)“ is the best way to go!
Apparently there are even papers where people train neural nets to fine-tune the parameters of neural nets, because of course there are 
1 „Gefällt mir“