REPAIR: Removing Representation Bias by Dataset Resampling

Li, Yang; Vasconcelos, Nuno

doi:10.1109/cvpr.2019.00980

Cited by 210 publications

(136 citation statements)

References 34 publications

Supporting

Mentioning

135

Contrasting

Order By: Relevance

“…the ability of simple linear classifiers to predict them correctly. While AFLite, among others (Li and Vasconcelos, 2019;Gururangan et al, 2018), advocate removing "easy" instances from the dataset, our work shows that easy-to-learn instances can be useful. Similar intuitions have guided other work such as curriculum learning (Bengio et al, 2009) and self-paced learning (Kumar et al, 2010;Lee and Grauman, 2011) where all examples are prioritized based on their "difficulty".…”

Section: Related Workmentioning

confidence: 85%

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Swayamdipta¹,

Schwartz²,

Lourie³

et al. 2020

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

168

135

View full text Add to dashboard Cite

Large datasets have become commonplace in NLP research. However, the increased emphasis on data quantity has made it challenging to assess the quality of data. We introduce Data Maps-a model-based tool to characterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for building data maps. This yields two intuitive measures for each example-the model's confidence in the true class, and the variability of this confidence across epochs-obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced characteristics. First, our data maps show the presence of ambiguous regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are easy to learn for the model, and play an important role in model optimization. Finally, data maps uncover a region with instances that the model finds hard to learn; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-ofdistribution generalization.

show abstract

Section: Related Workmentioning

confidence: 85%

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Swayamdipta¹,

Schwartz²,

Lourie³

et al. 2020

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

168

135

View full text Add to dashboard Cite

show abstract

“…This renders the above methods to account for confounders unsuitable as they either result in reduced number of training samples (e.g., matching or stratification) or require deterministic features that are computed beforehand (e.g., standardization or regression). Possible alternatives could be unbiased [17][18][19][20][21] and invariant feature-learning approaches [22][23][24][25] relying on end-to-end training to study the invariance (independence) between the learned feature F and a bias factor (① in Fig. 2b).…”

mentioning

confidence: 99%

Training confounder-free deep learning models for medical applications

2020

View full text Add to dashboard Cite

The presence of confounding effects (or biases) is one of the most critical challenges in using deep learning to advance discovery in medical imaging studies. Confounders affect the relationship between input data (e.g., brain MRIs) and output variables (e.g., diagnosis). Improper modeling of those relationships often results in spurious and biased associations. Traditional machine learning and statistical models minimize the impact of confounders by, for example, matching data sets, stratifying data, or residualizing imaging measurements. Alternative strategies are needed for state-of-the-art deep learning models that use end-to-end training to automatically extract informative features from large set of images. In this article, we introduce an end-to-end approach for deriving features invariant to confounding factors while accounting for intrinsic correlations between the confounder(s) and prediction outcome. The method does so by exploiting concepts from traditional statistical methods and recent fair machine learning schemes. We evaluate the method on predicting the diagnosis of HIV solely from Magnetic Resonance Images (MRIs), identifying morphological sex differences in adolescence from those of the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA), and determining the bone age from X-ray images of children. The results show that our method can accurately predict while reducing biases associated with confounders. The code is available at https://github.com/qingyuzhao/br-net.

show abstract

“…There are also debiasing strategies that identify hard examples in the dataset and re-train the model to focus on those examples (Yaghoobzadeh et al, 2019;Li and Vasconcelos, 2019;Le Bras et al, 2020). This approach reflects a similar in-tuition that simplicity is connected to dataset bias, although our method is able to explicitly model the bias and does not assume a pool of bias-free examples exist within the training data.…”

Section: Related Workmentioning

confidence: 99%

Learning to Model and Ignore Dataset Bias with Mixed Capacity Ensembles

Clark¹,

Yatskar²,

Zettlemoyer³

2020

Findings of the Association for Computational Linguistics: EMNLP 2020

View full text Add to dashboard Cite

Many datasets have been shown to contain incidental correlations created by idiosyncrasies in the data collection process. For example, sentence entailment datasets can have spurious word-class correlations if nearly all contradiction sentences contain the word "not", and image recognition datasets can have telltale object-background correlations if dogs are always indoors. In this paper, we propose a method that can automatically detect and ignore these kinds of dataset-specific patterns, which we call dataset biases. Our method trains a lower capacity model in an ensemble with a higher capacity model. During training, the lower capacity model learns to capture relatively shallow correlations, which we hypothesize are likely to reflect dataset bias. This frees the higher capacity model to focus on patterns that should generalize better. We ensure the models learn non-overlapping approaches by introducing a novel method to make them conditionally independent. Importantly, our approach does not require the bias to be known in advance. We evaluate performance on synthetic datasets, and four datasets built to penalize models that exploit known biases on textual entailment, visual question answering, and image recognition tasks. We show improvement in all settings, including a 10 point gain on the visual question answering dataset.

show abstract

REPAIR: Removing Representation Bias by Dataset Resampling

Cited by 210 publications

References 34 publications

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics

Training confounder-free deep learning models for medical applications

Learning to Model and Ignore Dataset Bias with Mixed Capacity Ensembles

Contact Info

Product

Resources

About