Robust Learning from Untrusted Sources

Authors: Nikola Konstantinov, Christoph Lampert

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform an extensive experimental evaluation of our algorithm and demonstrate its ability to learn from all available data, while successfully suppressing the effect of corrupted or irrelevant sources. [...] We perform two large sets of experiments, following the setup considered in our paper.
Researcher Affiliation Academia Nikola Konstantinov 1 Christoph H. Lampert 1 1Institute of Science and Technology, Klosterneuburg, Austria. Correspondence to: Nikola Konstantinov <nkonstan@ist.ac.at>.
Pseudocode Yes Pseudocode of the algorithm is given in Algorithm 1. Algorithm 1 Robust learning from untrusted sources Inputs: 1. Loss L, hypothesis set H, parameter λ 2. Reference dataset ST 3. Datasets S1, . . . , SN from the N sources for i = 1 to N do {Potentially in parallel} Compute d H (Si, ST ) end for Select α by solving (6). Minimize α-weighted loss: ˆhα = argminh Hˆϵα(h) Return: ˆhα
Open Source Code Yes Code is available at https://github.com/Nikola Kon1994/Robust Learning-from-Untrusted-Sources
Open Datasets Yes Our first set of experiments is on the "Multitask dataset of product reviews"2 (Pentina & Lampert, 2017), containing customer reviews for 957 Amazon products from the "Amazon product data" (Mc Auley et al., 2015a;b)... 2http://cvml.ist.ac.at/productreviews/. [...] The Animals with Attributes 2 dataset (Xian et al., 2018) contains 37322 images of 50 animal classes.
Dataset Splits Yes The hyperparameter λ is selected by 5-fold cross-validation on the trusted data. The regularization parameter is always selected by 5-fold cross-validation on the reference data.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments.
Software Dependencies No The paper mentions using a "pretrained model from the Tensor Nets package" but does not provide a version number for this or any other software dependencies (e.g., specific versions of Python, PyTorch, TensorFlow, or scikit-learn).
Experiment Setup Yes The hyperparameter λ is selected by 5-fold cross-validation on the trusted data. [...] All three methods use linear predictors and are trained by regularized logistic regression. The regularization parameter is always selected by 5-fold cross-validation on the reference data. [...] In our experiments, we use the recommended threshold value of c = 1.3452, under which the estimate of the linear predictor has been shown to achieve a 95% asymptotic relative efficiency (Pregibon, 1982).