Comparing the keras-to-loihi example notebook and the "equivalent" keras-to-snn example notebook from the NengoDL documentation, one of the differences that have not been documented is the use of use_bias=False in the NengoLoihi example.
The use of biases is disabled so that TF doesn't create sparse connections to implement the biases, since sparse connections is something NengoLoihi does not support. However, this point is overlooked (not mentioned) in the example notebook, and it would be good to stress this difference between the NengoLoihi and NengoDL examples for users comparing the two.
Comparing the keras-to-loihi example notebook and the "equivalent" keras-to-snn example notebook from the NengoDL documentation, one of the differences that have not been documented is the use of
use_bias=Falsein the NengoLoihi example.The use of biases is disabled so that TF doesn't create sparse connections to implement the biases, since sparse connections is something NengoLoihi does not support. However, this point is overlooked (not mentioned) in the example notebook, and it would be good to stress this difference between the NengoLoihi and NengoDL examples for users comparing the two.