Learning to Model and Ignore Dataset Bias with Mixed Capacity Ensembles

Abstract

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.

End-to-End Self-Debiasing Framework for Robust NLU Training

“…The results are satisfactory, especially when considering the simplicity and ef-ficiency of our approach. Moreover, the fact that a single configuration works well on 3 tasks is an indicator that our method has the potential to generalize on completely unknown OOD sets (Clark et al, 2020).…”

“…Following (Clark et al, 2019;Grand and Belinkov, 2019;Clark et al, 2020;Sanh et al, 2020), we tune our method hyperparameters on the OOD sets. As pointed out by (Clark et al, 2019(Clark et al, , 2020, this is not ideal since it assumes some prior knowledge of the OOD test sets. To best mitigate this impact, we follow the procedure of previous works and use the same hyper-parameters for all 3 tasks.…”

“…Recently, there has been a number of endeavours to produce a bias model without prior knowledge on the targeted biases or without the need for manually designing features. Utama et al (2020b) propose to use instead a model trained on a tiny fraction (< 1%) of the training data for few epochs as a bias model; while Clark et al (2020) and Sanh et al (2020) trained a low capacity model on the full training set. These approaches target the training of the bias model alone, which is subsequently queried while training the main model of interest.…”

Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition

“…Belinkov et al (2019) used adversarial training to mitigate the hypothesis-only bias in textual entailment models. Clark et al (2020) adversarially trained a low and a high capacity model in an ensemble in order to ensure that the latter model is focusing on patterns that should generalize better. Dayanik and Padó (2020) Dai and Adel (2020) explored different entity substitution techniques for data augmentation tailored to NER.…”

X-GGM: Graph Generative Modeling for Out-of-distribution Generalization in Visual Question Answering

“…To generalize to outof-distribution samples adaptively, the VQA model should own two capabilities: (1) overcoming negative language biases and (2) producing out-of-distribution answers by learning rules entailed in in-domain data. The prevailing OOD generalization methods [10,11,18,65] focus on enhancing the first capability, which achieves OOD generalization by explicitly mitigating the language biases. While the second capability, which directly endues VQA models the potentiality to generalize to out-of-distribution (i.e., unseen or rare) samples, has not been well explored.…”