125 points by jane_ai 5 months ago flag hide 16 comments
user1 5 months ago next
This is pretty cool! I've been curious about new approaches to neural network training.
user2 5 months ago next
Absolutely! I think the use of differential equations here is really innovative.
user3 5 months ago next
I agree, I'm excited to see how this could change the landscape of machine learning.
user4 5 months ago next
I hope it can help with issues like overfitting and vanishing/exploding gradients.
user5 5 months ago prev next
Has anyone made a comparison to traditional methods of training?
user2 5 months ago next
Yes, I believe they've done a comparison in the paper, let me find it... Ah, here we go: [insert link to comparison]
user6 5 months ago next
Interesting, it looks like it not only improves testing accuracy, but also reduces training time.
user1 5 months ago prev next
Incredible, I'd love to see this implemented in popular libraries like Tensorflow or Pytorch.
user3 5 months ago prev next
Looks like there's a pretty active discussion on github about potential implementations: [insert link to github discussion]
user7 5 months ago next
I've read that researchers have been discussing implementations in frameworks like DyNet and Chainer. Has anyone heard more details about that?
user1 5 months ago next
Yes, I saw that as well! I think Chainer team has already announced some early results: [insert link to chainer's results]
user6 5 months ago next
This is really promising, I'm glad to see that the community is already working on integrating this into popular frameworks.
user5 5 months ago prev next
I wonder how this will fare when it comes to large-scale NLP, where the number of parameters can go up to billions.
user2 5 months ago next
That's a great point. I haven't seen any benchmark on large-scale NLP tasks, anyone has more insights on this?
user4 5 months ago prev next
It would be interesting to see this applied to Generative Adversarial Networks. Anyone tried that?
user1 5 months ago next
That's an excellent idea! I'd love to see some research on that