Limitations of capsule networks
While capsule networks are great and they address the core issues of convolutional neural networks, they still have a long way to go. Some of the limitations of capsule networks are as follows:
- The network has not been tested on large datasets like ImageNet. This puts a question mark on their ability to perform well on large datasets.
- The algorithm is slow, mainly due to the inner loop of the dynamic routing algorithm. The number of iterations can be fairly large for large datasets.
- Capsule networks definitely have higher complexity in implementation compared to CNNs.
It would be interesting to see how the deep learning community addresses the limitations of capsule networks.
