Transfer and Marginalize: Explaining Away Label Noise with Privileged Information
Authors: Mark Collier, Rodolphe Jenatton, Effrosyni Kokiopoulou, Jesse Berent
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | TRAM performs strongly on CIFAR-10H, Image Net and Civil Comments benchmarks. and Empirically, we show that our method performs better than a series of baselines on CIFAR-10H, Image Net and Civil Comments benchmarks. |
| Researcher Affiliation | Industry | Mark Collier 1 Rodolphe Jenatton 1 EfiKokiopoulou 1 Jesse Berent 1 and 1Google AI. Correspondence to: Mark Collier <markcollier@google.com>. |
| Pseudocode | No | The paper describes the method using diagrams and textual explanations, but does not include any explicit pseudocode blocks or labeled algorithms. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code for the described methodology or provide a link to a code repository. |
| Open Datasets | Yes | We use both real-world datasets with PI as well as synthesizing PI for a re-labelled version of Image Net (Deng et al., 2009). and One dataset with annotator features is CIFAR-10H (Peterson et al., 2019), which is a re-labelled version of the CIFAR-10 (Krizhevsky & Hinton, 2009) test set. and Civil Comments1 is a collection of comments from independent news websites annotated with 7 toxicity labels... 1https://www.kaggle.com/c/jigsaw-unintended-bias-intoxicity-classification/data |
| Dataset Splits | Yes | The Identities subset consists of 405,130 training examples, 21,293 validation examples and 21,577 test set examples. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, or cloud computing instance specifications) used for running the experiments. |
| Software Dependencies | No | The paper mentions software like TensorFlow (via `tf.keras.applications`) and the Adam optimizer, but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | All models are trained for 20 epochs with the Adam optimizer with base learning rate= 0.001, β1 = 0.9, β2 = 0.999, ϵ = 1e 07. All models are trained with L2 weight regularization with weighting 1e 3. and All but Het-TRAM models are trained for 90 epochs with the SGD optimizer with base learning rate= 0.1, decayed by a factor of 10 after 30, 60 and 80 epochs. Following Collier et al. (2021), Het-TRAM is trained for 270 epochs with the same initial learning rate and learning rate decay at 90, 180 and 240 epochs. All models are trained with L2 weight regularization with weighting 1e 4. |