[visionlist] [source-code] A new type of layer for more robust Convolutional Networks

[Nicola Strisciuglio]

Dear all,
we released the code of the Push-Pull layer 
<https://github.com/nicstrisc/push-pull-cnn-layer>, which can be used as 
substitute of convolutional layers in ConvNet architectures, to increase 
their robustness and generalization to corruptions of the input images 
that have not been seen during training.

The code is available at https://github.com/nicstrisc/push-pull-cnn-layer
and refers at the paper 'Enhanced robustness of convolutional networks 
with a push–pull inhibition layer', which we published open access at
https://link.springer.com/article/10.1007/s00521-020-04751-8.

We hope this layer can be of use to make ConvNets more robust and to 
generalize better to unseen corruptions of the input.

All the best,
Nicola Strisciuglio

on behalf of
Manuel Lopez Antequera
Nicolai Petkov
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://visionscience.com/pipermail/visionlist_visionscience.com/attachments/20200510/37ee3a1e/attachment-0001.html>