Convolutional neural networks often generalize poorly tounseen domains . A novel approach is proposed based on probabilisticallymixing instance-level feature statistics of training samples across sourcedomains . Mixing styles of training instances results in novel domains being synthesized implicitly, which increase the diversity of the source domains, and hence the generalizability of the model . MixStyle fits into mini-batch training perfectly and isextremely easy to implement . The effectiveness of MixStyle is demonstrated on awide range of tasks including category classification, instance retrieval andreinforcement learning. MixStyle has been shown to be effective on a range of . tasks including . category . classification, . category classification and . instance retrieval . and .reinforcing learning .

Author(s) : Kaiyang Zhou, Yongxin Yang, Yu Qiao, Tao Xiang

Links : PDF - Abstract

Code :

https://github.com/alsoj/Recommenders-movielens


Coursera

Keywords : mixstyle - classification - training - instance - category -

Leave a Reply

Your email address will not be published. Required fields are marked *