Generalized Jensen-Shannon Divergence Loss for Learning with Noisy Labels
Authors: Erik Englesson, Hossein Azizpour
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide an extensive set of empirical evidences on several datasets, noise types and rates. They show state-of-the-art results and give in-depth studies of the proposed losses. [...] This section, first, empirically investigates the effectiveness of the proposed losses for learning with noisy labels on synthetic (Section 4.1) and real-world noise (Section 4.2). This is followed by several experiments and ablation studies (Section 4.3) to shed light on the properties of JS and GJS through empirical substantiation of the theories and claims provided in Section 2. |
| Researcher Affiliation | Academia | Erik Englesson KTH Stockholm, Sweden engless@kth.se Hossein Azizpour KTH Stockholm, Sweden azizpour@kth.se |
| Pseudocode | No | The paper describes its methods and algorithms in text and mathematical formulas but does not include explicit pseudocode blocks or algorithms labeled as such. |
| Open Source Code | Yes | implementation available at https://github.com/ErikEnglesson/GJS |
| Open Datasets | Yes | We evaluate the proposed loss functions on the CIFAR datasets with two types of synthetic noise: symmetric and asymmetric. [...] We use a smaller version called mini Web Vision [29] consisting of the first 50 classes of the Google subset. [...] In Appendix B.2, we show state-of-the-art results when using GJS on two other real-world noisy datasets: ANIMAL-10N [30] and Food-101N [31]. |
| Dataset Splits | Yes | We train and evaluate five networks for individual results, where in each run the synthetic noise, network initialization, and data-order are differently randomized. [...] We plot the evolution of the validation accuracy (a) and network s consistency (as measured by GJS) on clean (b) and noisy (c) examples of the training set of CIFAR-100 for varying symmetric noise rates when learning with the cross-entropy loss. |
| Hardware Specification | No | The paper does not specify the exact hardware (e.g., GPU models, CPU types, memory) used for the experiments. It mentions using 'ResNet 34 and 50 for experiments' but not the hardware on which these were run. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., 'PyTorch 1.x', 'Python 3.x'). It describes the experimental setup in Appendix A but does not list specific software versions. |
| Experiment Setup | Yes | The complete details of the training setup can be found in Appendix A. Most importantly, we take three main measures to ensure a fair and reliable comparison throughout the experiments: 1) we reimplement all the loss functions we compare with in a single shared learning setup, 2) we use the same hyperparameter optimization budget and mechanism for all the prior works and ours, and 3) we train and evaluate five networks for individual results, where in each run the synthetic noise, network initialization, and data-order are differently randomized. |