Agree to Disagree: Diversity through Disagreement for Better Transferability

Abstract

Gradient-based learning algorithms have an implicit simplicity bias which in effect can limit the diversity of predictors being sampled by the learning procedure. This behavior can hinder the transferability of trained models by (i) favoring the learning of simpler but spurious featurespresent in the training data but absent from the test data -and (ii) by only leveraging a small subset of predictive features. Such an effect is especially magnified when the test distribution does not exactly match the train distribution-referred to as the Out of Distribution (OOD) generalization problem. However, given only the training data, it is not always possible to apriori assess if a given feature is spurious or transferable.Instead, we advocate for learning an ensemble of models which capture a diverse set of predictive features. Towards this, we propose a new algorithm D-BAT (Diversity-By-disAgreement Training), which enforces agreement among the models on the training data, but disagreement on the OOD data. We show how D-BAT naturally emerges from the notion of generalized discrepancy, as well as demonstrate in multiple experiments how the proposed method can mitigate shortcut-learning, enhance uncertainty and OOD detection, as well as improve transferability.

Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations

“…Another group of papers proposes regularization techniques to learn diverse solutions on the train data, focusing on different groups of features [95,56,71,76]. Xu et al [101] show how to train orthogonal classifiers, i.e.…”

“…and Pagliardini et al[71], we consider Dominoes binary classification datasets, where the top half of the image shows MNIST digits[55] from classes {0, 1}, and the bottom half shows MNIST images from classes {7, 9} (MNIST-MNIST ), Fashion-MNIST[99] images from classes {coat, dress} (MNIST-Fashion) or CIFAR-10 [52] images from classes {car, truck} (MNIST-CIFAR). In all Dominoes datasets, the top half of the image (MNIST 0 − 1 images) presents a linearly separable feature; the bottom half of the image presents a harder to learn feature.…”

“… Shah et al [87] and Pagliardini et al[71] only considered Dominoes data where both the simple and complex features are perfectly predictive of the target, corresponding to a 100% spurious correlation.…”

Pathologies of Predictive Diversity in Deep Ensembles

“…While it is often assumed that the intuitions from tree-based ensembles carry over to this setting (e.g. Mishtal and Arel, 2012;Fort et al, 2019;Ross et al, 2020;Pagliardini et al, 2022), we find thatsurprisingly-deep ensemble performance is anticorrelated with predictive diversity. We demonstrate this phenomenon in Fig.…”

On Background Bias in Deep Metric Learning

“…Training an additional network that complements a biased model 19 or ensembles that learn diverse feature sets alleviate the problem that the model only learns a few potentially irrelevant features. 20 To prevent models from using spurious correlations between the image and the class label, 21-23 the network's last layer can be fine-tuned on data that does not show such correlations. 24 Kobs et al 17 investigate the influence of different image factors such as item or background color on different DML models.…”