Loss Function Learning for Domain Generalization by Implicit Gradient

Authors: Boyan Gao, Henry Gouk, Yongxin Yang, Timothy Hospedales

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the learned loss function on the four common DG benchmarks: VLCS (Fang et al., 2013), PACS (Li et al., 2018a), Office Home (Venkateswara et al., 2017), and Terra Incognita (Beery et al., 2018).
Researcher Affiliation Collaboration 1School of Informatics, University of Edinburgh 2Samsung AI Center, Cambridge.
Pseudocode Yes Further details can be found in (Lorraine et al., 2020), but we provide pseudo-code for computing the implicit gradient in Algorithm 2.
Open Source Code No The paper does not contain an explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes We choose Rotated MNIST (Ghifary et al., 2015) as a small dataset suitable for loss learning. We evaluate the learned loss function on the four common DG benchmarks: VLCS (Fang et al., 2013), PACS (Li et al., 2018a), Office Home (Venkateswara et al., 2017), and Terra Incognita (Beery et al., 2018).
Dataset Splits Yes We train Res Net-18 with ITL on the training split and perform model selection using the validation set.
Hardware Specification Yes training our ITL required using Py Torch (Paszke et al., 2017) only required 8 hours on a single V100 GPU to complete 200 gradient descent steps on ω.
Software Dependencies No The paper mentions using PyTorch and cites its creators, but does not provide a specific version number for PyTorch or any other software dependency.
Experiment Setup Yes The learning rates in the inner loop and outer loop are both 0.01, a batch size of 32 is used for the inner loop, and the Neumann series used for approximating the inverse Hessian is truncated at 15 iterations.