Benchmark: Layer-wise training of VGG-like network: 0.99% test error

[anokland]

Please consider adding this result to your list.

A VGG-like network with 6 convolutional layers and 1 fully connected layer.

conv128-conv256-maxpool-conv256-conv512-maxpool-conv512-maxpool-conv512-maxpool-fc1024-fc10

Standard preprocessing (mean/std subtraction/division)
Cutout data augmentation
7.3M parameters
Layer-wise training, no global back-propagation
Code and more results: https://github.com/anokland/local-loss

[tkasasagi]

Thank you so much. We will check it and update the list soon!

[mxbi]

Any chance you could rerun the model on the updated dataset and let us know what it achieves? I'll add it to the table then

[anokland]

Sure! With the updated dataset, the error dropped to 0.89%.