The weights from synapses saturated couple times

[Clea Smith]

Hi SpiNNaker group,

Recently I was trying to do unsupervised training on a local board. After extracting the weight, I tried to implement an inference net to check how the training accuracy is.

But I was given the following warning during inference. Just wondering what this means and does it matter to the simulation?

The weights from the synapses for IF_cond_exp_0:0:48 on 1, 1, 5 saturated 1 times. If this causes issues you can increase the spikes_per_second and / or ring_buffer_sigma values located within the .spynnaker.cfg file.
The weights from the synapses for IF_cond_exp_4:0:48 on 0, 0, 10 saturated 2 times. If this causes issues you can increase the spikes_per_second and / or ring_buffer_sigma values located within the .spynnaker.cfg file.

Many thanks

Clea :- >

[Andrew Rowley]

Hi,

 

These messages are due to the limited size of the buffers on SpiNNaker. 

 

To explain the process: we store the future weights to be added after a delay in ring-buffers, but the values are potentially summed over a number of time steps.  We attempt to work out what the maximum likely value in any one of these slots is using the statistics of the network, but this isn’t guaranteed to be an absolute maximum, only a likely maximum.  If enough spikes occur, the value in any ring buffer space might still therefore reach the maximum possible. 

 

If a maximum weight value is going to cause the neuron to spike anyway, the saturation of the value might not be considered critical.  Additionally, if it only happens a few times in the network run (in your case a total of 3 times), and the network is otherwise reasonably active, it is unlikely to cause any issues.

 

So my advice would be to continue to perform the tests, but be aware that this happened so if the analysis comes out looking wrong, this could be investigated further.  If the analysis proves to show that the check is fine, you can likely ignore the warnings.

 

Andrew :)

 

--
You received this message because you are subscribed to the Google Groups "SpiNNaker Users Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spinnakeruser...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/spinnakerusers/d7da5b22-2b79-4ab7-bad0-03b6d4fca3bfn%40googlegroups.com.