Can't fully define noise distribution

[jhtu]

The Reservoir constructor allows for various noise scalings (like noise_rc) and also the noise distribution (noise_type). However, I don't think there is anyway to further specify the noise. For instance, if we set noise_type='Uniform' and noise_rc=2. then we can draw values from $\text{Unif}(0, 2)$. However, if we want to instead draw values from $\text{Unif}(-1, 1)$, I don't see how we can do that.

The Reservoir class calls either base.forward_internal or base.forward_external, both of which use the reservoir's noise_generator, which in turn calls utils.random.noise. Actually, utils.random.noise accepts **kwargs and would allow for more flexible noise generation. But in base.forward_internal, there is no way to pass in other arguments.

[jhtu]

It looks like you can manually set the noise generator after calling the constructor, as so:

reservoir = rpy.nodes.Reservoir(...)
reservoir.hypers['noise_generator'] = partial(
    rpy.utils.random.noise, 
    seed=reservoir.hypers['noise_generator'].keywords['seed'], 
    low=-0.5, 
    high=0.5)

I believe this sets the noise generator to be a uniform distribution with values in [-0.5, 0.5).

[neuronalX]

Hello,

Could you be a bit more specific, where exactly do you want to add noise?
In the reservoir internal states? In the outputs (i.e. readout)? In the inputs? In some connections weights?

[jhtu]

I was trying to add noise in the reservoir hidden state update, as $c \xi$, but with $\xi \sim \text{Unif}(-1, 1)$.

[neuronalX]

Did you try deriving the reservoir node or the ESN node and adding the noise in the state update equation?

The noise generator you tried is from hyperopt library, which you probably don't need.

[nTrouvain]

Hello @jhtu,

A noise_kwargs argument is now available in Reservoir. Give it a dictionnary with keyword parameters from Numpy random generator distributions and it should work.

[jhtu]

Thanks for the quick fix. This works as expected!