https://twitter.com/hardmaru/status/1120863660107161600
Handwritten Character Recognition with Very Small Datasets
They show that capsule networks can generate new training samples from
existing ones, with realistic variations that are more human-like. They
get 98.7% on MNIST with only 200 training samples.
https://arxiv.org/abs/1904.08095
No comments:
Post a Comment