Searching for Robustness: Loss Learning for Noisy Classification Tasks

Abstract

We present a "learning to learn" approach for automatically constructing white-box classification loss functions that are robust to label noise in the training data. We parameterize a flexible family of loss functions using Taylor polynomials, and apply evolutionary strategies to search for noise-robust losses in this space. To learn re-usable loss functions that can apply to new tasks, our fitness function scores their performance in aggregate across a range of training dataset and architecture combinations. The resulting white-box loss provides a simple and fast "plug-andplay" module that enables effective noise-robust learning in diverse downstream tasks, without requiring a special training procedure or network architecture. The efficacy of our method is demonstrated on a variety of datasets with both synthetic and real label noise, where we compare favorably to previous work.

Fast and Efficient Local-Search for Genetic Programming Based Loss Function Learning

“…A promising alternative paradigm is to use evolution-based methods to learn M, favoring their inherent ability to avoid local optima via maintaining a population of solutions, their ease of parallelization of computation across multiple processors, and their ability to optimize for non-differentiable functions directly. Examples of such work include [19] and [23], which both represent M as parameterized Taylor polynomials optimized with covariance matrix adaptation evolutionary strategies (CMA-ES). These approaches generate interpretable loss functions, however; they also assume the parametric form via the degree of the polynomial.…”

Uncertainty quantification in skin cancer classification using three-way decision-based Bayesian deep learning

“…The weight of each individual classifier used at each step of the presented model can be considered as well [28]. Finally, both the robustness of the TWDBDL model against noise (noise classifications [47,20]) and the impact of various fusion models [5,39] can be studied.…”

Truncate-Split-Contrast: A Framework for Learning from Mislabeled Videos

“…Depending on whether noisy instances are detected in training, the existing LNL methods can be roughly divided into two types. One is to directly train a noise-robust model in the presence of noisy labels (Patrini et al 2017;Wang et al 2019;Ma et al 2020;Lyu and Tsang 2019;Zhou et al 2021;Gao, Gouk, and Hospedales 2021). The other one is to explicitly detect the potential noisy instances, and then learns a model by simply excluding them (Huang et al 2019), or re-using the potential noisy data by estimating the pseudo labels of them (Zhang et al 2018b;Li, Socher, and Hoi 2019;Li, Xiong, and Hoi 2021;Ortego et al 2021).…”