Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations
Authors: Yuli Slavutsky, Yuval Benjamini
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that our algorithm improves generalization to diverse class distributions in both simulations and experiments on real-world datasets. |
| Researcher Affiliation | Academia | Yuli Slavutsky Department of Statistics and Data Science The Hebrew University of Jerusalem Jerusalem, Israel yuli.slavutsky@mail.huji.ac.il Yuval Benjamini Department of Statistics and Data Science The Hebrew University of Jerusalem Jerusalem, Israel yuval.benjamini@mail.huji.ac.il |
| Pseudocode | Yes | Algorithm 1 Robust Zero-Shot Representation |
| Open Source Code | Yes | Code to reproduce our results is available at https://github.com/YuliSl/Zero_Shot_Robust_Representations. |
| Open Datasets | Yes | We used the ETHEC dataset [11] which contains 47,978 butterfly images... We used the Celeb A dataset [30] which contains 202,599 images of 10,177 celebrities. |
| Dataset Splits | No | The paper specifies class distributions for training and testing but does not provide explicit percentages or counts for a separate validation dataset split. It mentions that "hyperparameters for all methods were chosen via grid-search in a single experiment repetition" which implies a validation process, but the specific split ratios or methodology for a validation set are not detailed. |
| Hardware Specification | Yes | We ran all experiments on a single A100 cloud GPU. |
| Software Dependencies | Yes | All the code in this work was implemented in Python 3.10. We used the Tensor Flow 2.13 and Tensor Flow Addons 0.21 packages. For evaluation we used the auc function from scikitlearn 1.2. The Celeb A dataset was loaded through Tensor Flow Datasets 4.9 and pandas 1.5 was used to process the ETHEC dataset. Statistical tests were performed using ttest_rel and false_discovery_control functions from scipy.stats 1.11.4. All figures were generated using Matplotlib 3.7. |
| Experiment Setup | Yes | In all our experiments we used margin of m = 0.5 for the contrastive loss and Adam (Kingma & Ba, 2014) optimizer to train all models. ...hyperparameters for all methods were chosen via grid-search... All hyper-parameters are reported in Table 5. |