Unsupervised Domain Adaptation via Minimized Joint Error
Authors: Dexuan Zhang, Thomas Westfechtel, Tatsuya Harada
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide extensive empirical evidence showing that our proposal outperforms other upper bound related methods in image classification on several DA benchmarks, particularly when the domain gap is large. In this section, we evaluate our proposed method using several different datasets (Digit (Netzer et al., 2011; Le Cun et al., 1998; Hull, 1994), Vis DA (Peng et al., 2017), Office-Home (Venkateswara et al., 2017), and Office-31 (Saenko et al., 2010)). We conduct an ablation study (A.11) to demonstrate the contributions of each part of our proposal. |
| Researcher Affiliation | Academia | Dexuan Zhang EMAIL The University of Tokyo Thomas Westfechtel EMAIL The University of Tokyo Tatsuya Harada EMAIL The University of Tokyo RIKEN |
| Pseudocode | Yes | Algorithm 1 MJE |
| Open Source Code | No | The paper does not contain an explicit statement about releasing code for the methodology described, nor does it provide a direct link to a code repository. The Open Review link is for the review process, not for code. |
| Open Datasets | Yes | We evaluate our proposed method using several different datasets (Digit (Netzer et al., 2011; Le Cun et al., 1998; Hull, 1994), Vis DA (Peng et al., 2017), Office-Home (Venkateswara et al., 2017), and Office-31 (Saenko et al., 2010)). |
| Dataset Splits | Yes | SVHN MNIST We first examine the adaptation from SVHN (Fig. 7a) to MNIST (Fig. 7b). We used standard training and testing sets for both the source and target domains. MNIST USPS As for the adaptation between MNIST and USPS (Fig. 7c), we followed the training protocol established in Long et al. (2013) by sampling 2000 images from MNIST and 1800 from the USPS. For the test samples, the standard version was used for both source and target domains. |
| Hardware Specification | No | The paper mentions models like Res Net-101 and Res Net-50 and training processes, but it does not specify any particular hardware components such as GPU models, CPU models, or cloud computing instance types used for experiments. |
| Software Dependencies | No | The paper mentions several software components and algorithms used, such as "batch normalization", "spectral normalization (Miyato et al., 2018)", "Adam (Kingma & Ba, 2014)", and "Stochastic gradient descent (SDG) with Nesterov moment". However, it does not provide specific version numbers for any libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages (e.g., Python 3.x). |
| Experiment Setup | Yes | Throughout the experiment, we employed the CNN architecture used in Saito et al. (2017b), where batch normalization was applied to each layer, and a 0.5 rate of dropout was used between fully-connected layers. Adam (Kingma & Ba, 2014) was used for optimization, with a minibatch size of 128 and a learning rate of 10 4. Stochastic gradient descent (SDG) with Nesterov moment was used for optimization with a minibatch size of 32 and an initial learning rate of 10 3, which decayed exponentially. We tested γ = {0.1, 0.5, 0.9, 1} and η = {0, 0.5, 0.8, 0.9}. β = 0.01 was used in all the experiments. |