Open msseibel opened 5 years ago
Hi @msseibel , @abdel and I worked on a derived capsules algorithm which is described here:
https://agi.io/2018/11/29/predictive-capsules-networks-research-update/
We eventually tested MNIST, AffNIST, SmallNORB, and LFW (Labelled Faces in the Wild). The common feature of these datasets is that they all have a figure-ground layout - some thing in the middle, and a blank background without confusing distractors. Whereas, CIFAR has strongly textured backgrounds and relatively smaller objects. Without an attentional mechanism the Capsules routing might be unable to converge.
We didn't publish our results on the modified capsules algorithm because while it worked, it didn't generalize well. Have a read of the summary in the link and let us know if you want to know more.
Hello @abdel,
have you pursued your work on deeper capsule networks as stated in the conclusion of your paper? It seems like capsules can not deal with complex datasets like Cifar100. Why?
Are there too many irrelevant informations, which should rather not be reconstructed?
Finally, I see that Cifar10 is part of your input_data. Can you report results for it?
Thanks alot for your work. It is inspiring.