search

pumpikano / tf-dann

Domain-Adversarial Neural Network in Tensorflow
MIT License
628 stars 224 forks source link

Improving results for MNIST-M? #2

Closed zumpchke closed 7 years ago

zumpchke commented 7 years ago

I was able to reproduce the main result on the mnist-m dataset (basically~ 0.5 acc -> ~0.7). What are architectural strategies available to improve it even further?

pumpikano commented 7 years ago

I can refer you to two related works that offer interesting structural ideas.

Deep Reconstruction-Classification Networks for Unsupervised Domain Adaptation is directly comparable to this work. Towards Principled Unsupervised Learning looks at the idea of unsupervised domain alignment in their experiments. Both of these ideas have structural ideas that might help improve this task.

I hope this is helpful!

zumpchke commented 7 years ago

Thanks. BTW, I implemented this repo in Keras.

https://github.com/fchollet/keras/pull/4031