Mitigating Spurious Correlations in Patch-wise Tumor Classification on High-Resolution Multimodal Images

Patch-wise multi-label classification provides an efficient alternative to full pixel-wise segmentation on high-resolution images, particularly when the objective is to determine the presence or absence of target objects within a patch rather than their precise spatial extent. This formulation substantially reduces annotation cost, simplifies training, and allows flexible patch sizing aligned with the desired level of decision granularity. In this work, we focus on a special case, patch-wise binary classification, applied to the detection of a single class of interest (tumor) on high-resolution multimodal nonlinear microscopy images. We show that, although this simplified formulation enables efficient model development, it can introduce spurious correlations between patch composition and labels: tumor patches tend to contain larger tissue regions, whereas non-tumor patches often consist mostly of background with small tissue areas. We further quantify the bias in model predictions caused by this spurious correlation, and propose to use a debiasing strategy to mitigate its effect. Specifically, we apply GERNE, a debiasing method that can be adapted to maximize worst-group accuracy (WGA). Our results show an improvement in WGA by approximately 7% compared to ERM for two different thresholds used to binarize the spurious feature. This enhancement boosts model performance on critical minority cases, such as tumor patches with small tissues and non-tumor patches with large tissues, and underscores the importance of spurious correlation-aware learning in patch-wise classification problems.

Gradient Extrapolation for Debiased Representation Learning

Machine learning classification models trained with empirical risk minimization (ERM) often inadvertently rely on spurious correlations. When absent in the test data, these unintended associations between non-target attributes and target labels lead to poor generalization. This paper addresses this problem from a model optimization perspective and proposes a novel method, Gradient Extrapolation for Debiased Representation Learning (GERNE), designed to learn debiased representations in both known and unknown attribute training cases. GERNE uses two distinct batches with different amounts of spurious correlations to define the target gradient as the linear extrapolation of two gradients computed from each batch's loss. It is demonstrated that the extrapolated gradient, if directed toward the gradient of the batch with fewer amount of spurious correlation, can guide the training process toward learning a debiased model. GERNE can serve as a general framework for debiasing with methods, such as ERM, reweighting, and resampling, being shown as special cases. The theoretical upper and lower bounds of the extrapolation factor are derived to ensure convergence. By adjusting this factor, GERNE can be adapted to maximize the Group-Balanced Accuracy (GBA) or the Worst-Group Accuracy. The proposed approach is validated on five vision and one NLP benchmarks, demonstrating competitive and often superior performance compared to state-of-the-art baseline methods.

Fill in the Gap! Combining Self-supervised Representation Learning with Neural Audio Synthesis for Speech Inpainting

Most speech self-supervised learning (SSL) models are trained with a pretext task which consists in predicting missing parts of the input signal, either future segments (causal prediction) or segments masked anywhere within the input (non-causal prediction). Learned speech representations can then be efficiently transferred to downstream tasks (e.g., automatic speech or speaker recognition). In the present study, we investigate the use of a speech SSL model for speech inpainting, that is reconstructing a missing portion of a speech signal from its surrounding context, i.e., fulfilling a downstream task that is very similar to the pretext task. To that purpose, we combine an SSL encoder, namely HuBERT, with a neural vocoder, namely HiFiGAN, playing the role of a decoder. In particular, we propose two solutions to match the HuBERT output with the HiFiGAN input, by freezing one and fine-tuning the other, and vice versa. Performance of both approaches was assessed in single- and multi-speaker settings, for both informed and blind inpainting configurations (i.e., the position of the mask is known or unknown, respectively), with different objective metrics and a perceptual evaluation. Performances show that if both solutions allow to correctly reconstruct signal portions up to the size of 200ms (and even 400ms in some cases), fine-tuning the SSL encoder provides a more accurate signal reconstruction in the single-speaker setting case, while freezing it (and training the neural vocoder instead) is a better strategy when dealing with multi-speaker data.