Institution profile

Gran Sasso Science Institute

Academic institutioneurope · it
Official website
Research library85linked papers
Opportunities0open roles
Selected work

Representative Papers

Neural Rank Collapse: Weight Decay and Small Within-Class Variability Yield Low-Rank Bias

Feb 06, 2024arXiv.org

This work investigates the origin of low-rank bias in deep neural networks and its connection to neural collapse. For general feedforward networks with nonlinear activations, we propose the “neural rank collapse” mechanism: weight decay jointly with intra-class variance in hidden layers drives rapid singular value decay across weight matrices, inducing progressive rank reduction. We establish, for the first time in nonlinear deep networks, a quantitative theoretical link between low-rank bias and neural collapse—extending beyond existing linear-network analyses. Our theory proves that the rank decay rate is proportional to the intra-class variance of the preceding layer’s hidden representations. Using singular value analysis, statistical modeling of latent-space distributions, and extensive experiments across architectures (ResNet, CNN), we empirically validate the mechanism. Furthermore, leveraging this insight, we achieve controllable rank compression of weight matrices by over 30% without sacrificing accuracy.

8 citationsRead paper

Trigger Optimization and Event Classification for Dark Matter Searches in the CYGNO Experiment Using Machine Learning

Jan 28, 2026

This work addresses the challenges of efficient real-time triggering, compression, and background suppression in high-resolution sparse optical imaging within the CYGNO experiment. To this end, two novel approaches are proposed: first, an unsupervised online compression framework based on a convolutional autoencoder that extracts regions of interest (ROIs) via reconstruction residuals, achieving fully unsupervised real-time ROI identification for the first time in CYGNO—retaining 93.0% of signal intensity while discarding 97.8% of background pixels with only 25 ms inference latency; second, the application of the weakly supervised Classification Without Labels (CWoLa) method to mixed data, which successfully identifies compact, circular nuclear recoil events and achieves performance approaching the theoretical optimum.

1 citationsRead paper
Recent publications

Latest Papers