2
$\begingroup$

Capsule networks seem to match performance of convolutional neural networks on image classification tasks (more specifically on classification of handwritten digits in the MNIST dataset) 1.

I have also read in various articles that capsule networks require less training data than convolutional neural networks. Does this mean that a capsule network could reach the same accuracy as convolutional neural networks with less data? Eg. could a capsule network achieve the same accuracy as convolutional neural networks on the MNIST while only training on a fraction of the dataset?

$\endgroup$
1
$\begingroup$

Actually, the literature so far suggests that convolutional networks have better accuracy than capsule networks on many tasks, so it's by no means clear that capsule networks are superior across the board to CNNs. This is perhaps not surprising; we have many years of experience working out how to engineer convolutional networks so we get the best possible performance out of them, whereas capsule networks are still new and haven't yet benefitted from that degree of engineering. So, we might see further improvements in the performance of capsule networks in the coming years, now that they are getting more attention.

There are some specific areas where capsule networks have significantly outperformed convolutional networks. Those seem to be primarily those where viewpoint invariance is especially critical. There are others where so far they are worse. For instance, capsule networks so far have done significantly worse than convolutional networks on CIFAR10. There is active research and an active debate in the machine learning community about how well capsule networks will perform. I think it's too early to say how that will pan out.

As far as amount of training data, it is likely that this will depend on the task as well, and is probably an open problem at this point. I would imagine that capsule networks might need significantly less training data for tasks where viewpoint invariance is especially central, but that might not be true across all tasks.

In general, not a lot is known yet about the performance of capsule networks, and the answer to most questions of the form "how well will capsule networks perform on X?" is "you'll have to try it and see".

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.