Learning with Symmetric Label Noise: The Importance of Being Unhinged

Authors: Brendan van Rooyen, Aditya Menon, Robert C. Williamson

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now illustrate that the unhinged loss SLN-robustness is empirically manifest. We reiterate that with high regularisation, the unhinged solution is equivalent to an SVM (and in the limit any classification-calibrated loss) solution. Thus, we do not aim to assert that the unhinged loss is better than other losses, but rather, to demonstrate that its SLN-robustness is not purely theoretical. We first show that the unhinged risk minimiser performs well on the example of Long and Servedio [2010] (henceforth LS10). Figure 1 shows the distribution D, where X = {(1, 0), (γ, 5γ), (γ, γ)} R2, with marginal distribution M = { 1 2} and all three instances are deterministically positive. We pick γ = 1/2. The unhinged minimiser perfectly classifies all three points, regardless of the level of label noise (Figure 1). The hinge minimiser is perfect when there is no noise, but with even a small amount of noise, achieves a 50% error rate. We next consider empirical risk minimisers from a random training sample: we construct a training set of 800 instances, injected with varying levels of label noise, and evaluate classification performance on a test set of 1000 instances. We compare the hinge, t-logistic (for t = 2) [Ding and Vishwanathan, 2010] and unhinged minimisers using a linear scorer without a bias term, and regularisation strength λ = 10 16. From Table 1, even at 40% label noise, the unhinged classifier is able to find a perfect solution. By contrast, both other losses suffer at even moderate noise rates. We next report results on some UCI datasets, where we additionally tune a threshold so as to ensure the best training set 0-1 accuracy. Table 2 summarises results on a sample of four datasets. (The Appendix contains results with more datasets, performance metrics, and losses.) Even at noise close to 50%, the unhinged loss is often able to learn a classifier with some discriminative power.
Researcher Affiliation Collaboration Brendan van Rooyen , Aditya Krishna Menon , Robert C. Williamson , The Australian National University National ICT Australia { brendan.vanrooyen, aditya.menon, bob.williamson }@nicta.com.au
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide an unambiguous statement or link regarding the release of source code for the methodology described.
Open Datasets Yes We next report results on some UCI datasets, where we additionally tune a threshold so as to ensure the best training set 0-1 accuracy. Table 2 summarises results on a sample of four datasets. (The Appendix contains results with more datasets, performance metrics, and losses.)
Dataset Splits No The paper mentions training and testing sets, and tuning on the training set, but does not explicitly define a separate validation set or split percentages for validation.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers needed to replicate the experiments.
Experiment Setup Yes We compare the hinge, t-logistic (for t = 2) [Ding and Vishwanathan, 2010] and unhinged minimisers using a linear scorer without a bias term, and regularisation strength λ = 10 16. From Table 1, even at 40% label noise, the unhinged classifier is able to find a perfect solution. By contrast, both other losses suffer at even moderate noise rates. We next report results on some UCI datasets, where we additionally tune a threshold so as to ensure the best training set 0-1 accuracy.