Spuriosity Rankings: Sorting Data to Measure and Mitigate Biases
Authors: Mazda Moayeri, Wenxiao Wang, Sahil Singla, Soheil Feizi
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our method on Image Net, annotating 5000 classfeature dependencies (630 of which we find to be spurious) and generating a dataset of 325k soft segmentations for these features along the way. Having computed spuriosity rankings via the identified spurious neural features, we assess biases for 89 diverse models and find that class-wise biases are highly correlated across models. |
| Researcher Affiliation | Collaboration | Mazda Moayeri1 Wenxiao Wang1 Sahil Singla2,1 Soheil Feizi1 1 University of Maryland 2 Google {mmoayeri, wwx, sfeizi} @umd.edu, sasingla@google.com |
| Pseudocode | No | The paper describes procedures in narrative text and through diagrams, but it does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks with structured steps. |
| Open Source Code | Yes | We open source all code, and modify a key framework of our method to vastly increase its usability for people with fewer computational or data resources. |
| Open Datasets | Yes | We demonstrate our method on Image Net, annotating 5000 classfeature dependencies... We take UTKFace [60], a dataset of face images... We demonstrate this light-weight extension of the feature discovery framework of [46] on two spurious correlation benchmarks: Waterbirds, and Celeb-A Hair Color classification. |
| Dataset Splits | Yes | We use 90% of the data for training, and hold out the remaining 10%. (for UTKFace) ... Figure 6 (left) shows accuracy on the k = 10 lowest (y-axis) vs. highest (x-axis) spuriosity images for 89 diverse models pretrained on Image Net... Specifically, we obtain the training images from the class who s activation on the feature is within the top 20th percentile for the class. This results in a set of 260 (20% of 1300 training images per class) images per class-feature pair, as opposed to only 65 in [46]. |
| Hardware Specification | No | The paper mentions "This tuning occurs in minutes on a single GPU" but does not specify the model or type of GPU (e.g., NVIDIA A100, RTX 2080 Ti) or any other specific hardware components like CPU or memory. |
| Software Dependencies | No | The paper mentions using specific optimizers and model libraries (e.g., Adam optimizer, timm library), but it does not provide version numbers for any software dependencies (e.g., Python, PyTorch, specific library versions). |
| Experiment Setup | Yes | Specifically, we tune the classification heads of five models spanning diverse architectures (Res Net and Vi T) and training procedures (supervised, self-supervised, adversarial) on the 100 lowest spuriosity images for each of the 357 classes for which we discover one spurious feature. This tuning occurs in minutes on a single GPU, especially since the data passes through the entire model only once; after caching features, all computation is limited to the linear classification head. Also, we employ early stopping, halting tuning once spurious gap drops below 5%, so to avoid overfitting to the minority subpopulations at the cost of overall accuracy. ... In all cases, we use an Adam optimizer with learning rate of 0.1 and weight decay of 0.003. We train for 20 epochs. |