Identifying Spurious Biases Early in Training through the Lens of Simplicity Bias
This addresses the issue of spurious biases in neural networks for machine learning practitioners, offering a lightweight solution to improve robustness, though it is incremental as it builds on existing simplicity bias theory.
The paper tackles the problem of neural networks learning spurious correlations due to simplicity bias, showing that such features become separable early in training and can dominate predictions, leading to poor worst-group test accuracy. It proposes SPARE, a method that identifies spurious correlations early and uses importance sampling to mitigate them, achieving up to 21.1% improvement in worst-group accuracy and being up to 12x faster than state-of-the-art methods.
Neural networks trained with (stochastic) gradient descent have an inductive bias towards learning simpler solutions. This makes them highly prone to learning spurious correlations in the training data, that may not hold at test time. In this work, we provide the first theoretical analysis of the effect of simplicity bias on learning spurious correlations. Notably, we show that examples with spurious features are provably separable based on the model's output early in training. We further illustrate that if spurious features have a small enough noise-to-signal ratio, the network's output on the majority of examples is almost exclusively determined by the spurious features, leading to poor worst-group test accuracy. Finally, we propose SPARE, which identifies spurious correlations early in training and utilizes importance sampling to alleviate their effect. Empirically, we demonstrate that SPARE outperforms state-of-the-art methods by up to 21.1% in worst-group accuracy, while being up to 12x faster. We also show that SPARE is a highly effective but lightweight method to discover spurious correlations.