Learning advisor networks for noisy image classification

TL;DR

This work tackles the problem of training image classifiers with noisy labels by introducing Meta Feature Re-Weighting (MFRW), which uses an advisor network learned via meta-learning to adjust feature activations during training. The advisor learns to weight feature components of the backbone conditionally on the current sample loss, enabling the main model to exploit information in mislabeled data rather than discard it. Empirical results on CIFAR-10/100 with synthetic noise and Clothing1M show state-of-the-art performance, particularly under high noise, and demonstrate robust generalization. The approach highlights the potential of feature-level attention guided by meta-learning as a versatile strategy for learning under label noise and may extend to other problems with noisy annotations.

Abstract

In this paper, we introduced the novel concept of advisor network to address the problem of noisy labels in image classification. Deep neural networks (DNN) are prone to performance reduction and overfitting problems on training data with noisy annotations. Weighting loss methods aim to mitigate the influence of noisy labels during the training, completely removing their contribution. This discarding process prevents DNNs from learning wrong associations between images and their correct labels but reduces the amount of data used, especially when most of the samples have noisy labels. Differently, our method weighs the feature extracted directly from the classifier without altering the loss value of each data. The advisor helps to focus only on some part of the information present in mislabeled examples, allowing the classifier to leverage that data as well. We trained it with a meta-learning strategy so that it can adapt throughout the training of the main model. We tested our method on CIFAR10 and CIFAR100 with synthetic noise, and on Clothing1M which contains real-world noise, reporting state-of-the-art results.

Figures (1)

• Figure 1: Illustration of an iteration of the proposed Meta Feature Re-Weighting (MFRW) method. Each iteration is divided into four steps. First, a Loss Pre-Calculation is performed to calculate in advance the loss $\mathcal{L}^{pre}$ value of the training batch $x^{train}$. The second step is the Virtual-Train, where a clone of the main model is virtually updated. Here the meta-model modifies the feature of the main model by multiplying it with a vector of weights. The purple color indicates the weighted features. The third step shows the Meta-Train process. With a meta batch of clean example $x^{meta}$ the meta-model is updated minimizing the loss $\mathcal{L}^{meta}$ given by the previous virtually updated network. In the last phase Actual-Train, the real main model is trained with the meta-model optimized (yellow color)