Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization

• URL: http://arxiv.org/abs/2110.10832v1
• Date: Thu, 21 Oct 2021 00:08:17 GMT
• Title: Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization
• Authors: Devansh Arpit, Huan Wang, Yingbo Zhou, Caiming Xiong
• Abstract summary: In Domain Generalization (DG) settings, models trained on a given set of training domains have notoriously chaotic performance on shifted test domains. We first show that a simple protocol for averaging model parameters along the optimization path, starting early during training, significantly boosts domain generalizationity. We show that an ensemble of independently trained models also has a chaotic behavior in the DG setting.
• Score: 63.28279815753543
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: In Domain Generalization (DG) settings, models trained on a given set of training domains have notoriously chaotic performance on distribution shifted test domains, and stochasticity in optimization (e.g. seed) plays a big role. This makes deep learning models unreliable in real world settings. We first show that a simple protocol for averaging model parameters along the optimization path, starting early during training, both significantly boosts domain generalization and diminishes the impact of stochasticity by improving the rank correlation between the in-domain validation accuracy and out-domain test accuracy, which is crucial for reliable model selection. Next, we show that an ensemble of independently trained models also has a chaotic behavior in the DG setting. Taking advantage of our observation, we show that instead of ensembling unaveraged models, ensembling moving average models (EoA) from different runs does increase stability and further boosts performance. On the DomainBed benchmark, when using a ResNet-50 pre-trained on ImageNet, this ensemble of averages achieves $88.6\%$ on PACS, $79.1\%$ on VLCS, $72.5\%$ on OfficeHome, $52.3\%$ on TerraIncognita, and $47.4\%$ on DomainNet, an average of $68.0\%$, beating ERM (w/o model averaging) by $\sim 4\%$. We also evaluate a model that is pre-trained on a larger dataset, where we show EoA achieves an average accuracy of $72.7\%$, beating its corresponding ERM baseline by $5\%$.

Related papers

• DoGE: Domain Reweighting with Generalization Estimation [42.32000165235568]
  We propose DOmain reweighting with Generalization Estimation (DoGE) In our experiments, we extensively show how DoGE improves the generalization of the base model to any target data mixture. DoGE can effectively identify inter-domain dependencies, and consistently achieves better test perplexity on the target domain.
  arXiv  Detail & Related papers  (2023-10-23T22:51:58Z)
• Towards Unified and Effective Domain Generalization [42.579796989908914]
  We propose $textbfUniDG$, a novel and $textbfUnified framework for $textbfD$omain $textbfG$eneralization. Specifically, we encourage models to learn the distribution of test data in an unsupervised manner and impose a penalty regarding the updating step of model parameters.
  arXiv  Detail & Related papers  (2023-10-16T02:05:03Z)
• STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning [82.03481509373037]
  Recently, model-based reinforcement learning algorithms have demonstrated remarkable efficacy in visual input environments. We introduce Transformer-based wORld Model (STORM), an efficient world model architecture that combines strong modeling and generation capabilities. Storm achieves a mean human performance of $126.7%$ on the Atari $100$k benchmark, setting a new record among state-of-the-art methods.
  arXiv  Detail & Related papers  (2023-10-14T16:42:02Z)
• DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining [148.90031913522648]
  We propose Domain Reweighting with Minimax Optimization (DoReMi) DoReMi first trains a small proxy model using group distributionally robust optimization (Group DRO) over domains to produce domain weights. We then resample a dataset with these domain weights and train a larger, full-sized model.
  arXiv  Detail & Related papers  (2023-05-17T17:58:13Z)
• Improving Domain Generalization with Domain Relations [77.63345406973097]
  This paper focuses on domain shifts, which occur when the model is applied to new domains that are different from the ones it was trained on. We propose a new approach called D$3$G to learn domain-specific models. Our results show that D$3$G consistently outperforms state-of-the-art methods.
  arXiv  Detail & Related papers  (2023-02-06T08:11:16Z)
• ZooD: Exploiting Model Zoo for Out-of-Distribution Generalization [65.58562481279023]
  We propose ZooD, a paradigm for PTMs ranking and ensemble with feature selection. We evaluate our paradigm on a diverse model zoo consisting of 35 models for various Out-of-Distribution (OoD) tasks.
  arXiv  Detail & Related papers  (2022-10-17T16:31:57Z)
• Efficient Pre-trained Features and Recurrent Pseudo-Labeling inUnsupervised Domain Adaptation [6.942003070153651]
  We show how to efficiently opt for the best pre-trained features from seventeen well-known ImageNet models in unsupervised DA problems. We propose a recurrent pseudo-labeling model using the best pre-trained features (termed PRPL) to improve classification performance.
  arXiv  Detail & Related papers  (2021-04-27T21:35:28Z)
• Adaptive Risk Minimization: Learning to Adapt to Domain Shift [109.87561509436016]
  A fundamental assumption of most machine learning algorithms is that the training and test data are drawn from the same underlying distribution. In this work, we consider the problem setting of domain generalization, where the training data are structured into domains and there may be multiple test time shifts. We introduce the framework of adaptive risk minimization (ARM), in which models are directly optimized for effective adaptation to shift by learning to adapt on the training domains.
  arXiv  Detail & Related papers  (2020-07-06T17:59:30Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.