Auxiliary Task Reweighting for Minimum-data Learning
Authors: Baifeng Shi, Judy Hoffman, Kate Saenko, Trevor Darrell, Huijuan Xu
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In multiple experimental settings (e.g. semi-supervised learning, multi-label classification), we demonstrate that our algorithm can effectively utilize limited labeled data of the main task with the benefit of auxiliary tasks compared with previous task reweighting methods. We also show that under extreme cases with only a few extra examples (e.g. few-shot domain adaptation), our algorithm results in significant improvement over the baseline. |
| Researcher Affiliation | Collaboration | Baifeng Shi Peking University bfshi@pku.edu.cn Judy Hoffman Georgia Institute of Technology judy@gatech.edu Kate Saenko Boston University & MIT-IBM Watson AI Lab saenko@bu.edu Trevor Darrell, Huijuan Xu University of California, Berkeley {trevor, huijuan}@eecs.berkeley.edu |
| Pseudocode | Yes | Algorithm 1 ARML (Auxiliary Task Reweighting for Minimum-data Learning) |
| Open Source Code | Yes | Our code and video is available at https://sites.google.com/view/auxiliary-task-reweighting. |
| Open Datasets | Yes | Specifically, we use Self-supervised Semi-supervised Learning (S4L) [55] as our baseline algorithm. S4L uses self-supervised methods on unlabeled part of training data, and trains classifier on labeled data as normal. Following [55], we use two kinds of self-supervised methods: Rotation and Exemplar-MT. ... we test on two widely-used benchmarks: CIFAR-10 [27] with 4000 out of 45000 images labeled, and SVHN [40] with 1000 out of 65932 images labeled. ... We use the Celeb A dataset [34]. ... We use a common benchmark PACS [30] which contains four distinct domains of Photo, Art, Cartoon and Sketch. |
| Dataset Splits | No | The paper mentions percentages of labeled data used for the main task in semi-supervised learning and multi-label classification, and refers to 'test error', but it does not specify explicit validation dataset splits (e.g., percentages or sample counts for validation sets) needed for reproduction. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions high-level approaches and algorithms but does not specify software names with version numbers (e.g., PyTorch version, Python version, CUDA version) needed for replication. |
| Experiment Setup | No | The paper mentions aspects of the training process like using Langevin dynamics with Gaussian noise and adjusting learning rates, but it does not provide specific hyperparameter values (e.g., concrete learning rates, batch sizes, number of epochs) or a detailed experimental setup configuration. |