Optimizing Black-box Metrics with Iterative Example Weighting

Authors: Gaurush Hiranandani, Jatin Mathur, Harikrishna Narasimhan, Mahdi Milani Fard, Sanmi Koyejo

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various label noise, domain shift, and fair classification setups confirm that our proposal compares favorably to the state-of-the-art baselines for each application.
Researcher Affiliation Collaboration 1University of Illinois at Urbana-Champaign, Illinois, USA 2Google Research, USA 3Google Research, Accra.
Pseudocode Yes Algorithm 1: Elicit Weights for Diagonal Linear Metrics; Algorithm 2: Plug-in with Elicited Weights (PI-EW) for Diagonal Linear Metrics; Algorithm 3: Frank-Wolfe with Elicited Gradients (FW-EG) for General Diagonal Metrics (also depicted in Fig. 1)
Open Source Code Yes The source code (along with random seeds) is provided on the link below.1 https://github.com/koyejolab/fweg/
Open Datasets Yes We train a 10-class image classifier for the CIFAR-10 dataset (Krizhevsky et al., 2009); Our next experiment borrows the proxy label setup from Jiang et al. (2020) on the Adult dataset (Dua & Graff, 2017); The task is to learn a gender recognizer for the Adience face image dataset (Eidinger et al., 2014).
Dataset Splits Yes We take 2% of original training data as validation data and flip labels in the remaining training set...; We sample 1% validation data from the original training data...; For the validation set, we sample 20% of the 6 8 age bucket images.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU/CPU models or memory.
Software Dependencies No The paper mentions general software like SGD and ResNet, but does not specify exact version numbers for programming languages, libraries, or frameworks.
Experiment Setup Yes The learning rate for Fine-tuning is chosen from 1e{ 6,..., 4}. For PI-EW and FW-EG, we tune the parameter ϵ from {1, 0.4, 1e {4,3,2,1}}. The line search for Plug-in is performed with a spacing of 1e 4.