Learning with Bad Training Data via Iterative Trimmed Loss Minimization

Authors: Yanyao Shen, Sujay Sanghavi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we demonstrate its effectiveness in three settings: (a) deep image classifiers with errors only in labels, (b) generative adversarial networks with bad training images, and (c) deep image classifiers with adversarial (image, label) pairs (i.e., backdoor attacks). (Abstract) and 6. Experiments
Researcher Affiliation Academia 1ECE Department, University of Texas at Austin, TX, USA. Correspondence to: Yanyao Shen <shenyanyao@utexas.edu>, Sujay Sanghavi <sanghavi@mail.utexas.edu>.
Pseudocode Yes Algorithm 1 Iterative Trimmed Loss Minimization (ITLM) and Algorithm 2 Batch SGD Model Update(θ, S, t)
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes classification for CIFAR-10 with 40% random errors in labels (Figure 1 caption) and 5% subsampled MNIST (Le Cun et al., 1998) 3 dataset (Section 6.2) and CIFAR-10 (Krizhevsky & Hinton, 2009) (Section 6.2).
Dataset Splits No We demonstrate the effectiveness of ITLM for correcting training label errors in classification by starting from a clean dataset, and introducing either one of two different types of errors to make our training and validation data set (Section 6.2). However, specific split percentages or counts for training/validation/test are not provided.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running experiments.
Software Dependencies No The paper mentions types of neural networks and architectures (e.g., 'Wide Res Net-16', 'DC-GAN'), but does not specify any software libraries or their version numbers used for implementation.
Experiment Setup Yes We set α to be 5% less than the true ratio of clean samples, to simulate the robustness of our method to mis-specified sample ratio. (Section 6.2) and For the CIFAR-10 experiments, we run 4 rounds with early stopping, and then 4 rounds with full training. (Section 6.2) and We use ITLM with 4 early stopping rounds and 1 full training round, we set α as 0.98. (Section 6.4).