Asymmetric Multi-task Learning Based on Task Relatedness and Loss
Authors: Giwoong Lee, Eunho Yang, Sung Hwang
ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform experiments on multiple datasets for classification and regression, on which we obtain significant improvements in performance over the single task learning and existing multitask learning models. We validate our method on multiple datasets for classification and regression tasks, and obtain significant improvement over the single-task learning and exiting multi-task methods. Table 1 shows the prediction performance of the baselines and our methods on all four datasets. |
| Researcher Affiliation | Academia | Giwoong Lee SOPP0002@UNIST.AC.KR School of Electrical and Computer Engineering, UNIST, Ulsan, South Korea Eunho Yang EUNHOY@CS.KAIST.AC.KR School of Computing, KAIST, Daejon, South Korea Sung Ju Hwang SJHWANG@UNIST.AC.KR School of Electrical and Computer Engineering, UNIST, Ulsan, South Korea |
| Pseudocode | Yes | Algorithm 1 AMTL using Alternating Optimization and Algorithm 2 AMTL with Curriculum Learning |
| Open Source Code | No | The paper does not include an unambiguous statement about releasing code for the work described, nor does it provide a direct link to a source-code repository. |
| Open Datasets | Yes | 1) MNIST Digits data: This dataset contains 60, 000 training images and 10, 000 test images from 10 handwritten digits (0-9). 2) USPS Digits data: Another handwritten digit dataset, that is composed of 7, 291 training images and 2, 007 test images. 3) School dataset: This regression dataset consists of exam scores of 15,362 students from 139 schools, where the scores are real values. ...following the procedure of Argyriou et al. (2008). 4) AWA: This dataset (Lampert et al., 2009) contains 30, 475 images from 50 animal classes... Image Net-Room: This dataset is a subset of the Image Net dataset... |
| Dataset Splits | Yes | We generate total of 90 samples per task and use {30,30,30} split for training/validation/test. We use 5 random splits for training/validation/test datasets with 1000/500/500 instances, following the procedure of Kang et al. (2011) and Kumar & Daume III (2012) We generate 5 random splits by selecting 1000 random samples from the training set, and select two sets of 500 random images from the test set, to be used as validation and test. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software like 'weighted lasso implementation' and 'Caffe features' but does not provide specific version numbers for these or any other ancillary software dependencies. |
| Experiment Setup | No | The paper states that regularization parameters λ and µ are found through cross-validation, but it does not explicitly provide the specific values for these or other experimental setup details such as learning rates, batch sizes, or optimizer settings. |