A deep learning based radiomics model for differentiating intraparenchymal hematoma induced by cerebral venous thrombosis

A recent study published on arXiv presents a non-asymptotic convergence analysis of stochastic gradient Langevin dynamics (SGLD) in the lazy training regime, demonstrating that SGLD achieves exponential convergence to the empirical risk minimizer under certain conditions. The findings are supported by numerical examples in regression settings.

A novel method called Bayes-DIC Net has been introduced to estimate uncertainty in Digital Image Correlation (DIC) using Bayesian Neural Networks. This method generates high-quality datasets based on non-uniform B-spline surfaces, enabling the construction of realistic displacement fields for training deep learning algorithms in DIC applications.

The introduction of ImageNot, a dataset designed to be significantly different from ImageNet while maintaining a similar scale, reveals that deep learning models retain their ranking when evaluated on this new dataset. This finding suggests that the relative performance of models is consistent across different datasets, despite variations in absolute accuracy.

A recent study published on arXiv addresses the complexities of feature learning in deep learning, proposing a heuristic method for predicting the scales at which various patterns emerge. This approach simplifies the analytical challenges associated with high-dimensional non-linear equations often encountered in deep learning problems.

A recent publication on arXiv presents comprehensive lecture notes on regression analysis, aimed at students with basic university-level mathematics. The notes cover various regression techniques, including linear and logistic regression, and delve into advanced topics such as neural-network-based regression, providing a self-contained resource for understanding these methodologies.

A recent study has proposed a formalization of experimental studies in Machine Learning (ML) to better measure generalizability, addressing the challenge of ensuring that results can be replicated under varying conditions. This framework aims to quantify generalizability using rankings and Maximum Mean Discrepancy, providing insights into the necessary number of experiments for reliable outcomes.

Recent advancements in Spiking Neural Networks (SNNs) have led to the development of a novel training algorithm called S-SWIM, which adapts Random Feature Methods from Artificial Neural Networks. This approach allows for efficient training of SNNs without the need for approximating the spike function gradient, addressing a significant challenge in the field of machine learning.

A new method for assessing feature importance in Machine Learning, called Counterfactual Importance Distribution (CID), has been introduced. This post-hoc local feature importance method generates positive and negative counterfactuals, models their distributions using Kernel Density Estimation, and ranks features based on a distributional dissimilarity measure, enhancing the understanding of model decision-making processes.