Experiments (Dima)

MNIST experiments

Because of the poor quality on Office-31 dataset in the beginning, we decided to check if everything is correct on MNIST dataset. We have conducted 2 experiments with MNIST and MNIST-M domains and the results have showed that our model even slightly outperforms the original DANN. After that some experiments on SVHN and USPS domains were also made. The results are in the following table:

Source	Target	Train on target	Source only	Vanilla DANN	DANN w Alexnet	DANN paper
MNIST	MNIST-M	0.9715	0.4583	0.8652	0.7844	0.7666
MNIST-M	MNIST	0.9902	0.9673	0.9748
MNIST	SVHN	0.8823	0.2006	0.2601
SVHN	MNIST	0.9902	0.6489	0.6161	0.8353	0.7385
MNIST	USPS	0.9681	0.4036	0.7634
USPS	MNIST	0.9902	0.6225	0.9037	0.9410

Following parameters were used for training:

• batch_size = 32
• lr = 0.01
• n_epochs = 800
• domain_head = mnist_dann