interesting architecture
The reviewers have raised five issues: 1 Testing other architectures 2 The most interesting architectures, in our view, are locally connected but non-convolutional networks, like those in the
We will also investigate training AlexNet and VGG architectures using our approaches. We will show these plots in a new Appendix. In our ResNet-18 and ResNet-50 tests, the computational costs of WM's additional forward passes were 1.8 GFLOPs Could the brain have one-to-one wiring between forward and feedback neurons? We know that very precise and consistent neuronal wiring is found in simple organisms such as C. elegans And finally, something less than strict one-to-one wiring may suffice for effective learning.