AANG : Automating Auxiliary Learning

Authors: Lucio M. Dery, Paul Michel, Mikhail Khodak, Graham Neubig, Ameet Talwalkar

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental With natural language processing (NLP) as our domain of study, we demonstrate that our automated auxiliary learning pipeline leads to strong improvements over competitive baselines across continued training experiments on a pre-trained model on 5 NLP tasks 1.
Researcher Affiliation Collaboration Lucio M. Dery1 Paul Michel2 Mikhail Khodak1 Graham Neubig 1 Ameet Talwalkar1,3 1 Carnegie Mellon University 2 ENS PSL University 3 Hewlett Packard Enterprise
Pseudocode Yes Algorithm 1 AANG
Open Source Code Yes 1Code available at : https://github.com/ldery/Automating-Auxiliary-Learning.
Open Datasets Yes Table 4 in Appendix C provides details of the 5 datasets used. ... BIOMED CHEMPROT Kringelum et al. (2016) ... CS SCIERC Luan et al. (2018) ... STANCE SE-2016-6 Mohammad et al. (2016) ... CS ACL-ARC Jurgens et al. (2018) ... NEWS H.PARTISAN Kiesel et al. (2019)
Dataset Splits Yes Table 4: Specifications of datasets used to evaluate our methods. ... Train Size Dev Size Test Size
Hardware Specification Yes All models were trained on one of two types of gpus: NVIDIA A100 or NVIDIA A6000.
Software Dependencies No The paper mentions using "Adam W optimizer" and "Ro BERTabase" but does not provide specific version numbers for these or other software libraries/dependencies.
Experiment Setup Yes Training Details : Please see Appendix D for more details about hyper-parameter configurations. ... We use a batch size of 128 for all end-tasks tasks except H.PARTISAN where we use a batch size of 64. ... We use the Adam W optimizer (Loshchilov & Hutter, 2017), with weight decay of 0.01 for all experiments.