Co-regularized Alignment for Unsupervised Domain Adaptation

Authors: Abhishek Kumar, Prasanna Sattigeri, Kahini Wadhawan, Leonid Karlinsky, Rogerio Feris, Bill Freeman, Gregory Wornell

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed Co-regularized Domain Alignment (Co-DA) by instantiating it in the context of a recently proposed method VADA [35] which has shown state-of-the-art results on several benchmarks, and observe that Co-DA yields further significant improvement over it, establishing new state-of-the-art in several cases. For a fair comparison, we evaluate on the same datasets as used in [35] (i.e., MNIST, SVHN, MNIST-M, Synthetic Digits, CIFAR-10 and STL), and base our implementation on the code released by the authors3 to rule out incidental differences due to implementation specific details.
Researcher Affiliation Collaboration Abhishek Kumar MIT-IBM Watson AI Lab, IBM Research abhishk@us.ibm.com Prasanna Sattigeri MIT-IBM Watson AI Lab, IBM Research psattig@us.ibm.com Kahini Wadhawan MIT-IBM Watson AI Lab, IBM Research kahini.wadhawan@ibm.com Leonid Karlinsky MIT-IBM Watson AI Lab, IBM Research leonidka@il.ibm.com Rogerio Feris MIT-IBM Watson AI Lab, IBM Research rsferis@us.ibm.com William T. Freeman MIT billf@mit.edu Gregory Wornell MIT gww@mit.edu
Pseudocode No The paper describes the algorithmic instantiation of its approach in prose, but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper states 'base our implementation on the code released by the authors3' (referring to [35], VADA's authors), but does not explicitly state that the authors of *this* paper have released their own implementation code for Co-DA.
Open Datasets Yes For a fair comparison, we evaluate on the same datasets as used in [35] (i.e., MNIST, SVHN, MNIST-M, Synthetic Digits, CIFAR-10 and STL)
Dataset Splits Yes The hyperparameters are tuned by randomly selecting 1000 target labeled examples from the training set and using that for validation, following [35, 32].
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or cloud computing instances used for running the experiments.
Software Dependencies No The paper mentions 'Adam Optimizer' and 'batch-norm' and 'dropout' layers as part of the model and training process, but does not list any specific software dependencies or libraries with their version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For VADA hyperparameters λce and λsv (Eq. 7), we fix their values to what were reported by Shu et al. [35]... For the domain alignment hyperparameter λd, we do our own search over the grid {10 3, 10 2}... The hyperparameter for target prediction agreement, λp, was obtained by a search over {10 3, 10 2, 10 1}. For hyperparameters in the diversity term, we fix λdiv = 10 2 and do a grid search for ν (Eq. 5) over {1, 5, 10, 100}. The hyperparameters are tuned by randomly selecting 1000 target labeled examples from the training set and using that for validation, following [35, 32]. We completely follow [35] for training our model, using Adam Optimizer (lr = 0.001, β1 = 0.5, β2 = 0.999) with Polyak averaging (i.e., an exponential moving average with momentum= 0.998 on the parameter trajectory), and train the models in all experiments for 80k iterations as in [35].