Self-Paced Multitask Learning with Shared Knowledge
Authors: Keerthiram Murugesan, Jaime Carbonell
IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Results show that in each of the above formulations self-paced (easier-to-harder) task selection outperforms the baseline version of these methods in all the experiments. |
| Researcher Affiliation | Academia | Keerthiram Murugesan and Jaime Carbonell Carnegie Mellon University, Pittsburgh, PA, USA {kmuruges,jgc}@cs.cmu.edu |
| Pseudocode | Yes | The pseudo-code is in Algorithm 1. Algorithm 1: Self-Paced Multitask Learning: A General Framework |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | London School data (school) consists of examination scores of 15, 362 students from 139 schools in London. Each school is considered as a task and the feature set includes year of the examination, four school-specific and three studentspecific features. We replace each categorical feature with one binary variable for each possible feature value, as suggested in [Argyriou et al., 2008]. This results in 26 features with additional feature to account for the bias term. We use the ten 20% 80% train-test splits that came with the dataset for our experiments. Our proposed self-paced multitask learning algorithm does exceptionally better in school, which is a benchmark dataset for multitask experiments in the existing literature [Agarwal et al., 2010; Kumar and Daume, 2012]. |
| Dataset Splits | Yes | Unless otherwise specified, all model parameters are chosen via 3-fold cross validation. |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers. |
| Experiment Setup | Yes | All reported results in this section are averaged over 10 random runs of the training data. Unless otherwise specified, all model parameters are chosen via 3-fold cross validation. For all the experiments, we update the τ values using the equation 8. We use the ten 20% 80% train-test splits that came with the dataset for our experiments. Train-test splits are obtained by selecting 75% 25%, thus giving 15 examples for training and 5 examples for test set. We use 120 reviews per task for training and the rest of the reviews for test set. We use 80 examples from each task for training and the rest as the test data. We repeat the experiments on 10 (stratified) splits to measure the performance reliably. |