Identifiability Conditions for Domain Adaptation

Authors: Ishaan Gulrajani, Tatsunori Hashimoto

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present experiments validating our two algorithmic contributions, CERTIFY and WOMP. The focus of our work is conceptual rather than empirical; as such, our experiments consider illustrative simplified settings rather than realistic benchmarks.
Researcher Affiliation Academia Ishaan Gulrajani 1 Tatsunori B. Hashimoto 1 1Stanford University, USA.
Pseudocode Yes Algorithm 1 CERTIFY
Open Source Code No The paper does not provide any specific link to a code repository or an explicit statement about releasing the source code for the described methodology.
Open Datasets Yes We use CERTIFY to certify the linear asymmetry of MNIST at a confidence level of 95%. Our first experiment considers the standard MNIST-USPS benchmark task (Long et al., 2013), in which MNIST digits form the source domain and USPS digits form the target domain. The first is a semi-synthetic Colored MNIST dataset obtained by coloring half of the images in MNIST red and the other half green, yielding source and target domains. The second is a subset of the popular Domain Net benchmark (?), constructed by taking the real photos domain as the source and the Quick Draw drawings domain as the target.
Dataset Splits Yes first, we split each dataset into 50% training, 25% validation, and 25% test splits.
Hardware Specification No No specific hardware details (like GPU/CPU models, processor types, or memory amounts) used for running experiments are mentioned in the paper.
Software Dependencies No The paper mentions optimizers like Adam and uses specific architectural details (e.g., ReLU MLP), but does not provide specific version numbers for software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Training proceeds for 40K steps with encoder learning rates 10 3, classifier learning rate 10 3, and discriminator learning rate 10 3. We use a discriminator gradient penalty weight of 10. For MNIST-USPS experiments only, we use ℓ2 regularization with λ = 10 3 in the classifier and discriminator. MLPs have a single hidden layer of width 64. The CNNs have two layers of 5 5 kernels with stride 2 and widths 16 and 32.