Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation

Authors: Floris Holstege, Bram Wouters, Noud Van Giersbergen, Cees Diks

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

Reproducibility Variable Result LLM Response
Research Type Experimental By evaluating the algorithm on benchmark datasets from computer vision (Waterbirds, Celeb A) and natural language processing (Multi NLI), we show it outperforms existing concept-removal methods in terms of identifying the main-task and spurious concepts, while removing only the latter.
Researcher Affiliation Academia 1University of Amsterdam, Department of Quantitative Economics 2Tinbergen Institute. Correspondence to: Floris Holstege <f.g.holstege@uva.nl>.
Pseudocode Yes Algorithm 1 JSE algorithm to estimate orthonormal bases for Zsp and Zmt. The conditions in the if-statements are discussed in Section 3.3. Input: a sample {ymt,k, ysp,k, zk}n k=1 consisting of two binary labels and a vector zk Rd. Initialize embedding matrix Z = (z1 z2 zn) . Initialize Z sp Z. for i = 1, ..., d do Zremain Z sp for j = 1, ..., d do Estimate ˆwsp, ˆwmt with Equation 2 (use Zremain).
Open Source Code Yes Our code with an implementation of JSE is publicly available.* *https://github.com/fholstege/JSE
Open Datasets Yes Waterbirds: this dataset from Sagawa et al. (2020b) is a combination of the Places dataset (Zhou et al., 2016) and the CUB dataset (Welinder et al., 2010)... Celeb A: this dataset contains images of celebrity faces (Liu et al., 2015)... Multi NLI: the Multi NLI dataset (Williams et al., 2018)...
Dataset Splits Yes For a given dataset size (e.g. n =2,000) the data is split into an 80% training and 20% validation set, and a test set of the same size is kept apart for evaluation.
Hardware Specification No The paper specifies the neural network architectures (Res Net50, BERT) and software libraries (torchvision, transformers) used, but does not provide any concrete details about the specific hardware (e.g., GPU models, CPU types) on which experiments were run.
Software Dependencies No The paper mentions 'torchvision package', 'transformers package', 'Pytorch', and 'Adam optimizer' but does not specify exact version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes For waterbirds, this means using a learning rate of 10 3, a weight decay of 10 3, a batch size of 32, and for 100 epochs without early stopping. For Celeb A, this means using a learning rate of 10 3, a weight decay of 10 4, a batch size of 128, and for 50 epochs without early stopping.