Implementation is not the same as in paper about deepmiml HOT 3 CLOSED

The main propose of this repo is to show how to build a customer layer such as sub-concept layers. ( It might be a little tricky for beginners to build such thing) The activations for the cub-concept layers, on the other hand, is quite flexible.
For instance, ReLUs/Sigmoid/LinearMappings can be easily modified in the framework provided in this repo and it will be quite easy to create a ReLU sub-concept layer.

from deepmiml.

Thanks, did you try different implementation and how did it influence the sub-concept layer?

from deepmiml.

Haven't tried all but basically they all produce a matching score (linear, sigmiod, relu or even tanh) and thus they are all reasonable choices, performance might be different depending on the datasets.

from deepmiml.