Learning to Generalize: Meta-Learning for Domain Generalization
Authors: Da Li, Yongxin Yang, Yi-Zhe Song, Timothy Hospedales
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method and achieve state of the art results on a recent cross-domain image classification benchmark, as well demonstrating its potential on two classic reinforcement learning tasks. Experiments To evaluate our method, we compare it with various alternatives on four different problems, including an illustrative synthetic experiment, a challenging recent computer vision benchmark for multi-class classification across different domains, and two classic reinforcement learning problems, Cart-Pole and Mountain Car. |
| Researcher Affiliation | Academia | Da Li, Yongxin Yang, Yi-Zhe Song Queen Mary University of London {da.li, yongxin.yang, yizhe.song}@qmul.ac.uk Timothy M. Hospedales The University of Edinburgh t.hospedales@ed.ac.uk |
| Pseudocode | Yes | Algorithm 1 Meta-Learning Domain Generalization, Algorithm 2 MLDG for Reinforcement Learning |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the described methodology. |
| Open Datasets | Yes | Specifically, we used the PACS multi-domain recognition benchmark, a new dataset designed for the cross-domain recognition problems (Li et al. 2017)1. This dataset has 9991 images in total across 7 categories ( dog , elephant , giraffe , guitar , house , horse and person ) and 4 domains of different stylistic depictions ( Photo , Art painting , Cartoon and Sketch ). The footnote 1 links to http://sketchx.eecs.qmul.ac.uk. |
| Dataset Splits | Yes | For final-test, we use the best performing model on the validation set after 45k iterations. In each mini-batch, we split the S = 6 source domains into V = 2 meta-test and S V = 4 meta-train domains. This is to mimic real train-test domain-shifts so that over many iter- ations we can train a model to achieve good generalization in the final-test evaluated on target domains T. |
| 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 states 'We implement MLDG in TensorFlow' but does not specify the version number for TensorFlow or any other software dependencies. |
| Experiment Setup | Yes | We use SGD optimizer with learning rate 5e 4 (exponential decay is used with decay step 15k and decay rate 0.96) and mini-batch 64. Meanwhile, parameters α, β, γ are set to 5e 4, 1.0 and 5e 4. |