Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
This provides a simpler, more efficient solution for improving model robustness to spurious correlations, benefiting practitioners in computer vision and related fields.
The paper tackles the problem of neural networks relying on spurious features like backgrounds for predictions, showing they often still learn core features, and demonstrates that simple last layer retraining matches or outperforms state-of-the-art methods on spurious correlation benchmarks with much lower complexity and computational cost.
Neural network classifiers can largely rely on simple spurious features, such as backgrounds, to make predictions. However, even in these cases, we show that they still often learn core features associated with the desired attributes of the data, contrary to recent findings. Inspired by this insight, we demonstrate that simple last layer retraining can match or outperform state-of-the-art approaches on spurious correlation benchmarks, but with profoundly lower complexity and computational expenses. Moreover, we show that last layer retraining on large ImageNet-trained models can also significantly reduce reliance on background and texture information, improving robustness to covariate shift, after only minutes of training on a single GPU.