Mitigating Spurious Bias with Last-Layer Selective Activation Retraining
Keywords: spurious correlation, robustness, classification
Abstract: Deep neural networks trained with standard empirical risk minimization (ERM)
tend to exploit the spurious correlations between non-essential features and classes
for predictions. For example, models might identify an object using its frequently
co-occurring background, leading to poor performance on data lacking the correlation. Last-layer retraining approaches the problem of over-reliance on spurious correlations by adjusting the weights of the final classification layer. The success
of this technique provides an appealing alternative to the problem by focusing on
the improper weighting on neuron activations developed during training. However,
annotations on spurious correlations are needed to guide the weight adjustment. In
this paper, for the first time, we demonstrate theoretically that neuron activations,
coupled with their final prediction outcomes, provide self-identifying information
on whether the neurons are affected by spurious bias. Using this information,
we propose last-layer selective activation retraining (LaSAR), which retrains the
last classification layer while selectively blocking neurons that are identified as
spurious. In this way, we promote the model to discover robust decision rules
beyond spurious correlations. Our method works in a classic ERM training set-
ting where no additional annotations beyond class labels are available, making
it a practical and efficient post-hoc tool for improving a model’s robustness to
spurious correlations. We theoretically show that LaSAR brings a model closer to
the unbiased one and empirically demonstrate that our method is effective with
different model architectures and can effectively mitigate spurious bias on different
data modalities without requiring annotations of spurious correlations in data.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10275
Loading