Tilted Empirical Risk Minimization

Authors: Tian Li, Ahmad Beirami, Maziar Sanjabi, Virginia Smith

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we showcase the flexibility, wide applicability, and competitive performance of the TERM framework through empirical results on a variety of real-world problems such as handling outliers (Section 5.1), ensuring fairness and improving generalization (Section 5.2), and addressing compound issues (Section 5.3).
Researcher Affiliation Collaboration Tian Li CMU tianli@cmu.edu Ahmad Beirami Facebook AI beirami@fb.com Maziar Sanjabi Facebook AI maziars@fb.com Virginia Smith CMU smithv@cmu.edu
Pseudocode Yes Algorithm 1: Batch TERM", "Algorithm 2: Stochastic TERM", "Algorithm 3: Batch Non-Hierarchical TERM", "Algorithm 4: Stochastic Non-Hierarchical TERM
Open Source Code Yes We provide implementation details in Appendix J. All code, datasets, and experiments are publicly available at github.com/litian96/TERM.
Open Datasets Yes For regression tasks, we use the drug discovery data extracted from (Diakonikolas et al., 2019) which is originally curated from (Olier et al., 2018) and train linear regression models with different losses... For mitigating noise on classification tasks, we use the standard CIFAR-10 data and their standard train/val/test partitions along with a standard inception network (Szegedy et al., 2016)... We use the HIV-1 dataset (Rögnvaldsson, 2013) as in (Namkoong & Duchi, 2017)
Dataset Splits Yes We randomly split the dataset into 80% training set, 10% validation set, and 10% testing set.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) needed to replicate the experiment.
Experiment Setup Yes For experiments with positive t (Section 5.2), we tune t P t0.1, 0.5, 1, 5, 10, 50, 100, 200u on the validation set... For experiments regarding focal loss (Lin et al., 2017), we select the class balancing parameter (α in the original focal loss paper) from rangep0.05, 0.95, 0.05q and select the main parameter γ from t0.5, 1, 2, 3, 4, 5u... The initial step-size is set to 0.1 and decayed to 0.01 at epoch 50. The batch size is 100.