PES implementation in NengoDL

Backends
1

I’m trying to understand why the PES implementation in NengoDL requires redefining the build function [link].
This seems to be done to “avoid slicing encoders along an axis > 0” but I’m not sure what’s meant by this or why it would be a technical problem.
What’s the problem with doing the equivalent to error = dot(encoders, error) inside SimPESBuilder?

2

The build_pes function in Nengo Core has encoder slicing along axis > 0 at line 726 in learning_rules.py. Nengo-DL does not support this slicing along an axis > 0 and therefore it is required to have a custom build function.

The error = dot(encoders, error) is effectively the same.

3

Thanks!

What does that slicing actually do? In my topology:

pre = nengo.Ensemble( n_neurons, dimensions=dimensions, seed=seed )
post = nengo.Ensemble( n_neurons, dimensions=dimensions, seed=seed )
conn = nengo.Connection(
        pre.neurons,
        post.neurons,
        transform=np.zeros( (post.n_neurons, pre.n_neurons) )
        )

I see that conn.post_slice is always None, so local_error=error without any padding.

When would the problematic slicing come into effect?

4

It would come into effect if there is slicing of the post synaptic object like the following

conn = nengo.Connection(
        pre.neurons,
        post.neurons[:5],
        transform=np.zeros( (5, pre.n_neurons) )
        )

ther would be a not None conn.post_slice.