Searching for Robustness: Loss Learning for Noisy Classification Tasks

Abstract

We present a "learning to learn" approach for discovering white-box classification loss functions that are robust to label noise in the training data. We parameterise 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 datasets and architectures. The resulting white-box loss provides a simple and fast "plug-and-play" module that enables effective labelnoise-robust learning in diverse downstream tasks, without requiring a special training procedure or network architecture. The efficacy of our loss is demonstrated on a variety of datasets with both synthetic and real label noise, where we compare favourably to prior work.

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

“…For example, in [24,28], differentiable surrogates of non-differentiable performance metrics are learned to reduce the misalignment problem between the performance metric and the loss function. Alternatively, in [4,9,27,46], loss functions are learned to improve sample efficiency and asymptotic performance in supervised and reinforcement learning, while in [3,20,35], they improved on the robustness of a model to domain-shifts and improved domain-generalization.…”

Meta Mirror Descent: Optimiser Learning for Fast Convergence

“…MetaReg Balaji et al (2018) meta-learns regularization parameters to improve domain generalisation. ARL (Gao et al, 2021) meta-learns a loss function to improve robustness of learning form noisy labels.…”

“…This tells us that the expected convergence rate on novel tasks depends on the learning divergence on training tasks, plus complexity terms such as the F-norm of the meta-learned optimiser weights M . Note that restricting the diameter r of the parameter space is usually required to obtain generalisation guarantees (Bartlett et al, 2017;Long & Sedghi, 2020;Gouk et al, 2021), so this is not an unusual or counterproductive requirement.…”

Learning Symbolic Model-Agnostic Loss Functions via Meta-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 [16] and [17], which both represent M as parameterized Taylor polynomials optimized with covariance matrix adaptation evolutionary strategies (CMA-ES). These approaches successfully derive interpretable loss functions, but similar to previously, they also assume the parametric form via the degree of the polynomial.…”

“…In particular, many loss function learning approaches use a parametric loss function representation such as a neural network [15] or Taylor polynomial [16], [17], which is limited as it imposes unnecessary assumptions and constraints on the structure of the learned loss function. However, the current non-parametric alternative to this is to use a two-stage discovery and optimization process, which infers both the loss function structure and parameters simultaneously using genetic programming and covariance matrix adaptation [18], and quickly become intractable for large-scale optimization problems.…”