Unsupervised Domain Adaptation for Distance Metric Learning
Authors: Kihyuk Sohn, Wenling Shang, Xiang Yu, Manmohan Chandraker
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we first illustrate how FTN works in a controlled setting of adapting from MNIST-M to MNIST with disjoint digit classes between the two domains and then demonstrate the effectiveness of FTNs through state-of-the-art performances on a cross-ethnicity face recognition problem. |
| Researcher Affiliation | Collaboration | Kihyuk Sohn1 Wenling Shang2 Xiang Yu1 Manmohan Chandraker1,3 1NEC Labs America 2University of Amsterdam 3UC San Diego |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement or link for open-source code related to the described methodology. |
| Open Datasets | Yes | In experiments, we first perform on a controlled setting by adapting between disjoint sets of digit classes. Specifically, we adapt from 0 4 of MNIST-M (Ganin et al., 2016) dataset to 5 9 of MNIST dataset and demonstrate the effectiveness of FTN in learning to align and separate domains. Then, we assess the impact of our proposed unsupervised CD2MA method on a challenging cross-ethnicity face recognition task, whose source domain contains face images of Caucasian identities and the target domain of non-Caucasian identities, such as African-American or East-Asian. This is an important problem since existing face recognition datasets show significant label biases towards Caucasian ethnicity, leading to sub-optimal recognition performance for other ethnicities. The proposed method demonstrates significant improvement in face verification and identification compared to a source-only baseline model and a standard DANN. Our proposed method also closely matches the performance upper bounds obtained by training with fully labeled source and target domains. |
| Dataset Splits | Yes | For verification, following the standard protocol (Huang et al., 2007), we construct 10 splits, each containing 900 positive and 900 negative pairs, and compute the accuracy on each split using the threshold found from the other 9 splits. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used to run its experiments, such as CPU or GPU models. |
| Software Dependencies | No | The paper mentions 'Adam stochastic optimizer' but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | We use Adam stochastic optimizer with learning rate of 0.0003, λ1 = 0.3 and λ2 = 0.03 to train FTN. All models, including supervised CNNs (Sup C, Sup C,A,E), are trained with 4096-pair loss. For Sup C and Sup C,A,E, we use Adam stochastic optimizer (Kingma & Ba, 2015) with the learning rate of 0.0003 for the first 12K updates and 0.0001 and 0.00003 for the next two subsequent 3K updates. All modules are then updated with the learning rate of 0.00003. Hyperparameters of different models are summarized in Table S3. |