Mitigating Spurious Correlations via Disagreement Probability

Authors: Hyeonggeun Han, Sehwan Kim, Hyungjun Joo, Sangwoo Hong, Jungwoo Lee

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations on multiple benchmarks demonstrate that DPR achieves state-of-the-art performance over existing baselines that do not use bias labels.
Researcher Affiliation Collaboration 1ECE & 2Next Quantum, Seoul National University 3Hodoo AI Labs {hygnhan, sehwankim, joohj911, tkddn0606, junglee}@snu.ac.kr
Pseudocode Yes Algorithm 1: Disagreement Probability based Resampling for debiasing (DPR)
Open Source Code Yes Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes]
Open Datasets Yes Colored MNIST (C-MNIST) is a synthetic dataset designed for digit classification, comprising ten digits, each spuriously correlated with a specific color. Following the protocols in Ahn et al. [1], we set the ratios of bias-conflicting samples, denoted as ρ, at 0.5%, 1%, and 5% for the training set, and 90% for the unbiased test set.
Dataset Splits Yes Additionally, we use a 10% of training data as validation data, and an unbiased test set with a bias-conflicting ratio of 90% is employed for performance evaluation.
Hardware Specification Yes All classification models are trained using an NVIDIA RTX A6000.
Software Dependencies No The paper mentions optimizers (SGD, Adam, Adam W) and models (BERT) but does not provide specific version numbers for any software dependencies (e.g., library or framework versions like PyTorch 1.9).
Experiment Setup Yes We train the model for 100 epochs with SGD optimizer, a batch size of 128, a learning rate of 0.02, weight decay of 0.001, momentum of 0.9, and learning rate decay of 0.1 at every 40 steps.