On the Unreasonable Effectiveness of Last-layer Retraining
Last-layer retraining (LLR) methods -- wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training -- have garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this ``unreasonable effectiveness'' of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in the implicit bias of gradient descent benefiting robustness. Our empirical investigation does not support this hypothesis. Instead, we present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set. We conclude by showing how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Spuriosity Rankings for Free: A Simple Framework for Last Layer Retraining Based on Object Detection
Deep neural networks have exhibited remarkable performance in various domains. However, the reliance of these models on spurious features has raised concerns about their reliability. A promising solution to this problem …
Objectobject-detectionObject DetectionOpen-vocabulary object detection+1The Unreasonable Ineffectiveness of the Deeper Layers
We empirically study a simple layer-pruning strategy for popular families of open-weight pretrained LLMs, finding minimal degradation of performance on different question-answering benchmarks until after a large fraction…
GPUQuantizationQuestion AnsweringFew-shot Image Generation with Elastic Weight Consolidation
Few-shot image generation seeks to generate more data of a given domain, with only few available training examples. As it is unreasonable to expect to fully infer the distribution from just a few observations (e.g., emoj…
10-shot image generationDiversityImage GenerationFourier networks for uncertainty estimates and out-of-distribution detection
A simple method for obtaining uncertainty estimates for Neural Network classifiers (e.g. for out-of-distribution detection) is to use an ensemble of independently trained networks and average the softmax outputs. While t…
Out-of-Distribution DetectionUnreasonable Effectiveness of Last Hidden Layer Activations for Adversarial Robustness
In standard Deep Neural Network (DNN) based classifiers, the general convention is to omit the activation function in the last (output) layer and directly apply the softmax function on the logits to get the probability s…
Adversarial Robustness