Domain Separation Networks
Authors: Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, Dumitru Erhan
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our novel architecture results in a model that outperforms the state-of-the-art on a range of unsupervised domain adaptation scenarios and additionally produces visualizations of the private and shared representations enabling interpretation of the domain adaptation process. Experimental results and discussion are given in Sec. 4. We evaluate the efficacy of our method with each of the two similarity losses outlined in Sec. 3.2 by comparing against the prevailing visual domain adaptation techniques for neural networks: Correlation Alignment (CORAL) [26], Domain-Adversarial Neural Networks (DANN) [7, 8], and MMD regularization [29, 17]. |
| Researcher Affiliation | Collaboration | Konstantinos Bousmalis Google Brain Mountain View, CA konstantinos@google.com George Trigeorgis Imperial College London London, UK g.trigeorgis@imperial.ac.uk Nathan Silberman Google Research New York, NY nsilberman@google.com Dilip Krishnan Google Research Cambridge, MA dilipkay@google.com Dumitru Erhan Google Brain Mountain View, CA dumitru@google.com |
| Pseudocode | No | The paper describes the architecture and learning process using equations and descriptive text, but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | Yes | We provide code at https://github.com/tensorflow/models/domain_adaptation. |
| Open Datasets | Yes | We evaluate on object classification datasets used in previous work4 including MNIST and MNIST-M [8], the German Traffic Signs Recognition Benchmark (GTSRB) [25], and the Streetview House Numbers (SVHN) [20]. |
| Dataset Splits | Yes | Out of the 59, 001 MNIST-M training examples, we used the labels for 1, 000 of them to find optimal hyperparameters for our models. |
| Hardware Specification | No | The paper states models were implemented using Tensor Flow and trained with Stochastic Gradient Descent, but does not provide specific hardware details such as GPU/CPU models or types of machines used. |
| Software Dependencies | No | The paper mentions 'Tensor Flow [1]' but does not provide a specific version number for it or any other software dependencies. |
| Experiment Setup | Yes | Our initial learning rate was multiplied by 0.9 every 20, 000 steps (mini-batches). We used batches of 32 samples from each domain for a total of 64 and the input images were mean-centered and rescaled to [ 1, 1]. In order to avoid distractions for the main classification task during the early stages of the training procedure, we activate any additional domain adaptation loss after 10, 000 steps of training. |