Robust Classification via Regression for Learning with Noisy Labels
Authors: Erik Englesson, Hossein Azizpour
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments, that increases our understanding of the method and shows its strong performance compared to baselines on several datasets (Section 4). |
| Researcher Affiliation | Academia | Erik Englesson, Hossein Azizpour KTH Royal Institute of Technology {engless, azizpour}@kth.se |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at: github.com/Erik Englesson/SGN. |
| Open Datasets | Yes | We conduct experiments on the CIFAR-N (Wei et al., 2021b), Clothing1M (Xiao et al., 2015), and (mini) Web Vision Li et al. (2017) datasets. |
| Dataset Splits | Yes | For the experiments on the CIFAR (including CIFAR-N) datasets, we implement all baselines in the same shared code base to have an as conclusive comparison as possible. To achieve the best possible performance in this setup, we do a search for method-specific hyperparameters for each method based on a noisy validation set. |
| Hardware Specification | No | All experiments were performed using the supercomputing resource Berzelius provided by the National Supercomputer Centre at Linköping University and the Knut and Alice Wallenberg foundation. |
| Software Dependencies | No | The paper mentions using "Tensor Flow Probability (Dillon et al., 2017)" but does not specify a version number for this or any other software dependency. |
| Experiment Setup | Yes | All methods use the same Wide Res Net (WRN-28-2) architecture, with a constant learning rate (0.01), SGD with momentum (0.9) and weight decay (5e-4), batch size of 128, and standard data augmentation (crop and flip). We used 300 training epochs, but found that the baselines that estimate shifts/labels, ELR (Liu et al., 2020), SOP (Liu et al., 2022), NAL (Lu et al., 2022), and ours, benefited from more training epochs and were trained for twice as long. |