Central Moment Discrepancy (CMD) for Domain-Invariant Representation Learning
Authors: Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, Susanne Saminger-Platz
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test our approach on two different benchmark data sets for object recognition (Office) and sentiment analysis of product reviews (Amazon reviews). CMD achieves a new state-of-the-art performance on most domain adaptation tasks of Office and outperforms networks trained with Maximum Mean Discrepancy, Variational Fair Autoencoders and Domain Adversarial Neural Networks on Amazon reviews. In addition, a post-hoc parameter sensitivity analysis shows that the new approach is stable w. r. t. parameter changes in a certain interval. Our experimental evaluations are based on two benchmark datasets for domain adaptation, Amazon reviews and Office, described in subsection Datasets. The experimental setup is discussed in subsection Experimental Setup and our classification accuracy results are discussed in subsection Results. Subsection Parameter Sensitivity analysis the accuracy sensitivity w. r. t. parameter changes of K for CMD and β for MMD. |
| Researcher Affiliation | Collaboration | Werner Zellinger, Edwin Lughofer & Susanne Saminger-Platz Department of Knowledge-Based Mathematical Systems Johannes Kepler University Linz, Austria {werner.zellinger, edwin.lughofer, susanne.saminger-platz}@jku.at Thomas Grubinger & Thomas Natschl ager Data Analysis Systems Software Competence Center Hagenberg, Austria {thomas.grubinger, thomas.natschlaeger}@scch.at |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code of the experiments is publicly available1. 1https://github.com/wzell/cmd |
| Open Datasets | Yes | Amazon reviews: For our first experiment we use the Amazon reviews data set with the same preprocessing as used by Chen et al. (2012); Ganin et al. (2016); Louizos et al. (2016). Office: The second experiment is based on the computer vision classification data set from Saenko et al. (2010). |
| Dataset Splits | No | The paper mentions 'validation procedures' in relation to DANN, but does not explicitly provide details on validation dataset splits for their own experiments (e.g., percentages or counts for a validation set). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions software like 'Keras (Chollet, 2015)', 'Adagrad optimizer (Duchi et al., 2011)', and 'Adadelta optimizer from Zeiler (2012)', but it does not specify version numbers for these libraries or frameworks (e.g., Keras 2.x, TensorFlow 1.x). |
| Experiment Setup | Yes | For the CMD regularizer, the λ parameter of eq. (2) is set to 1, i.e. the weighting parameter λ is neglected. The parameter K is heuristically set to five... For the MMD regularizer we use the Gaussian kernel with parameter β. We performed a hyper-parameter search for β and λ... the parameter λ is tuned on 10 values between 0.1 and 500 on a logarithmic scale. The parameter β is tuned on 10 values between 0.01 and 10 on a logarithmic scale. We use a similar architecture as Ganin et al. (2016) with one dense hidden layer with 50 hidden nodes, sigmoid activation functions and softmax output function. For all evaluations, the default parametrization is used as implemented in Keras (Chollet, 2015). |