pyESN icon indicating copy to clipboard operation
pyESN copied to clipboard

Leaking Rate (a)

Open SpyroRock opened this issue 4 years ago • 4 comments

Is there anywhere in the code the leaking parameter (a) in order to calculate the states of the network?

Thank you.

SpyroRock avatar Nov 18 '19 13:11 SpyroRock

Hi, I'm not sure I parsed your question correctly, but there's currently no leakage parameter. In other words, the only way to control how 'excitable' the network is (i.e. how strongly and for how long it will reverberate in response to input) is by adjusting the magnitude of the recurrent weights and of the input (i.e. playing with spectral_radius and input_scaling). For example, with a lower spectral radius, activity will die out sooner.

cknd avatar Nov 18 '19 13:11 cknd

Thank you for your response, I think you answered my question!

SpyroRock avatar Nov 18 '19 14:11 SpyroRock

So, if I will try to add the leakage parameter, it would be pissible or the structure of the code doesn't allow such modification?
The spectral radius concerns the initialization of recurrent weight. The leaking rate concerns how quickly the network forgets the previous states. Right?

SpyroRock avatar Nov 18 '19 15:11 SpyroRock

yes, that should be possible! be sure to use a default value for the new parameter that keeps the current behavior (= no decay)

cknd avatar Nov 18 '19 16:11 cknd