SearcharxivSearch

arXiv subjects

Sebastian Popescu

Publications and source records attributed to Sebastian Popescu.

4 recordsLinked to original sources

NECO: NEural Collapse Based Out-of-distribution detection

Detecting out-of-distribution (OOD) data is a critical challenge in machine learning due to model overconfidence, often without awareness of their epistemological limits. We hypothesize that ``neural collapse'', a phenomenon affecting in-distribution data for models trained beyond loss convergence, also influences OOD data. To benefit from this interplay, we introduce NECO, a novel post-hoc method for OOD detection, which leverages the geometric properties of ``neural collapse'' and of principal component spaces to identify OOD data. Our extensive experiments demonstrate that NECO achieves state-of-the-art results on both small and large-scale OOD detection tasks while exhibiting strong generalization capabilities across different network architectures. Furthermore, we provide a theoretical explanation for the effectiveness of our method in OOD detection. Code is available at https://gitlab.com/drti/neco

stat.ML

Hierarchical Gaussian Processes with Wasserstein-2 Kernels

Stacking Gaussian Processes severely diminishes the model's ability to detect outliers, which when combined with non-zero mean functions, further extrapolates low non-parametric variance to low training data density regions. We propose a hybrid kernel inspired from Varifold theory, operating in both Euclidean and Wasserstein space. We posit that directly taking into account the variance in the computation of Wasserstein-2 distances is of key importance towards maintaining outlier status throughout the hierarchy. We show improved performance on medium and large scale datasets and enhanced out-of-distribution detection on both toy and real data.

stat.ML

Ball Lightning as a Self-organized Complexity

The ball lightning phenomenon is explained in the frame of a self-organization scenario suggested by experiments performed on the spontaneously generated complex spherical space charge configuration in plasma. Originated in a hot plasma, suddenly created in a point where a lightning flash strikes the Earth surface, the ball lightning appearance proves the ability of nature to generate, by self-organization, complex structures able to ensure their own existence by exchange of matter and energy with the surroundings. Their subsequent evolution depends on the environment where they are born. Under contemporary Earth conditions the lifetime of such complexities is relatively short. A similar self-organization mechanism, produced by simple sparks in the earl Earth's atmosphere (chemically reactive plasma) is suggested to explain the genesis of complexities able to evolve into prebiotic structures.

nlin.PS

On the "mystery" of differential negative resistance

Investigating the causes of the nonlinear behavior of a gaseous conductor we identified the presence of two states of the complex space charge configuration self-assembled in front of the anode. These states correspond to two levels of self-organization from which the first one is related to spatial pattern whereas the second one to spatiotemporal pattern. Their emergence through two kinds of instabilities produced for two critical distances from thermodynamic equilibrium is emphasized in the current voltage characteristic as an S-shaped, respectively Z-shaped bistability. Their presence attributes to the gaseous conductor the ability to work as an S-shaped, respectively an N-shaped negative differential resistance.

physics.plasm-ph