Uniform Convergence of Rank-weighted Learning

Authors: Justin Khim, Liu Leqi, Adarsh Prasad, Pradeep Ravikumar

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6. Experiments, Following our theoretical analysis, we test our results on logistic regression and linear regression. Our results for logistic regression are given in Figure 2, and our results for linear regression are given in Figure 3.
Researcher Affiliation Academia Justin Khim 1 Liu Leqi 1 Adarsh Prasad 1 Pradeep Ravikumar 1 1Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA. Correspondence to: Justin Khim <jkhim@cs.cmu.edu>.
Pseudocode No The paper describes an 'iterative optimization procedure' with an equation, but it is not presented as a formal pseudocode block or algorithm.
Open Source Code Yes The code of the experiments can be found at https://bit.ly/2YzwRk J.
Open Datasets No In the logistic regression setup, the features are drawn from X Uniform(B(d)) where B(d) is the d-dimension ball. The labels are sampled from a distribution over { 1, 1} where Y takes the value +1 with probability (1+X T *)/2. where * = (1, 0, ..., 0) Rd. In the linear regression experiment, we draw our covariates from a Gaussian X N(0, Id) in Rd. The noise distribution is fixed as N(0, 0.01). We draw our response variable Y as, Y= X T * + where * = (1, 1, ..., 1) Rd.
Dataset Splits No The paper describes generating samples for empirical evaluation and mentions sample sizes, but does not specify standard training/validation/test dataset splits or their percentages/counts.
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide any specific software dependencies with version numbers.
Experiment Setup Yes In the logistic regression setup, the features are drawn from X Uniform(B(d)) where B(d) is the d-dimension ball. The labels are sampled from a distribution over { 1, 1} where Y takes the value +1 with probability (1+X T *)/2. where * = (1, 0, ..., 0) Rd. We use the logistic loss ( ; (X, Y)) = log ( 1 + exp( YX T ) ) . For HRM, we have chosen a=0.6 and b=0.4. For CVa R and trimmed mean, is set to be 0.1. ... iterative optimization procedure to obtain a heuristic minimizer: t+1 = t - t / n * sum_i=1^n wt_i * ( t; Zi), (7) for all t in {0, ..., T-1}, where wt_i = w(Fn((t; Zi))) is the rank-dependent weighting and t is the learning rate.