Auxiliary Learning with Joint Task and Data Scheduling
Authors: Hong Chen, Xin Wang, Chaoyu Guan, Yue Liu, Wenwu Zhu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our JTDS model significantly outperforms the state-of-the-art methods under supervised, semisupervised and corrupted label settings. |
| Researcher Affiliation | Collaboration | Hong Chen 1 Xin Wang 1 2 Chaoyu Guan 1 Yue Liu 1 Wenwu Zhu 1 1Department of Computer Science and Technology, Tsinghua University 2THU-Bosch JCML center, Tsinghua University. |
| Pseudocode | Yes | Algorithm 1 presents the complete algorithm that simultaneously learns the task learning model and the task-data scheduler. |
| Open Source Code | Yes | Our code will be released at https://github.com/forchchch/JTDS |
| Open Datasets | Yes | Task and Dataset We first evaluate our methods on the fine-grained bird classification problem on the CUB (Wah et al., 2011) dataset... We conduct this experiment on CIFAR10, CIFAR100 (Krizhevsky & Hinton, 2009) and Oxford-IIIT Pet dataset (Parkhi et al., 2012)... on the Movie Lens-1M dataset (Harper & Konstan, 2016) |
| Dataset Splits | Yes | Let Dt = {(xt i, yt i1, yt i2, , yt in, yt i G)}m i=1 be the training set... Let Dv = {(xv j, yv j G)}N j=1 be the validation set... The datasets are split into training, validation and test set as the officially recommended. We sample 200 samples from the validation set to form Ddev. |
| Hardware Specification | No | The paper does not explicitly state the specific hardware used for running the experiments (e.g., GPU models, CPU models, or cloud instance types). It mentions using ResNet18 and other models, but these are software architectures, not hardware. |
| Software Dependencies | No | The paper mentions using SGD and Adam optimizers and models like ResNet18 and AutoINT, but it does not specify version numbers for any software dependencies, such as deep learning frameworks (e.g., PyTorch, TensorFlow) or their versions. |
| Experiment Setup | Yes | In the CIFAR100, CUB and Pet datasets, we use the Res Net18 (He et al., 2016) as the backbone... For the CIFAR10 dataset, we adopt a 4-layer Conv Net as the backbone... For the recommendation task, Auto INT (Song et al., 2019) is the adopted backbone... we optimize θ using SGD with initial learning rate searched from {0.01,0.02,0.05} and a cosine annealing scheduler, and in the CUB, Pet, Movie Lens-1M dataset we use Adam (Kingma & Ba, 2014) with learning rate searched from {1e-3,1e-4}. To optimize β, we use SGD with learning rate from {1e-2,1e-3}. Additionally, M in the Algorithm is searched from {10, 20, 50} and K is searched from {3,5}... The running epoch is 200 for CIFAR100 and Pet, 100 for CUB, 15 for Movie Lens-1M. |