A new arXiv paper introduces FICAug, a method that combines feature-informed clustering with data augmentation to improve facial-expression-based screening for Parkinson's disease. The approach targets the problem of small clinical datasets, which limits how well such screening models generalize. It is presented as an updated preprint on arXiv (2409.17685v3) in the cs.AI and cs.LG categories.
A theoretical paper examines the foundations of contrastive learning, a method that uses data augmentation to learn feature representations without large labeled datasets. The authors analyze how the structure of the augmentation graph relates to whether neural networks can approximate the resulting objective. The work aims to fill gaps in the theoretical understanding of why contrastive learning works in practice.
Researchers propose a conditional quantum flow matching approach for generating synthetic physiological signals when labeled data is scarce. Unlike earlier quantum generative models that begin from uninformative noise, the method incorporates class structure already present in the data. The work targets label-scarce physiological signal classification tasks.
Researchers propose a speech synthesis method that separates speaker identity from dysarthric articulation patterns, allowing finer control over generated dysarthric speech. The approach conditions synthesis on individual patients, producing varied synthetic speakers to supplement scarce training data for dysarthric speech recognition. This addresses a field bottlenecked by high speaker variability and limited labeled recordings.
A new arXiv paper introduces a unified pipeline that uses a single phoneme-based text-to-speech model to generate synthetic training data for automatic speech recognition. The work presents a controlled study of how text selection, reference speech, and augmentation scale affect recognition accuracy.