Publications

Satellite Image Time Series (SITS) data has proven effective for agricultural tasks due to its rich spectral and temporal nature. In this study, we tackle the task of stress detection in sugar-beet fields using a fully unsupervised approach. We propose a 3D convolutional autoencoder model to extract meaningful features from Sentinel-2 image sequences, combined with acquisition-date-specific temporal encodings to better capture the growth dynamics of sugar-beets. The learned representations are used in a downstream clustering task to separate stressed from healthy fields. The resulting stress detection system can be directly applied to data from different years, offering a practical and accessible tool for stress detection in sugar-beets.

The research extends diffusion model guidance to work with standard classifiers not trained on noise, unlike previous methods requiring robust classifiers. After analyzing classifier sensitivity on multiple datasets, researchers found non-robust classifiers produce unstable gradients under noise. Their solution uses one-step denoised predictions with stochastic optimization techniques like exponential moving averages, resulting in stable guidance that maintains sample diversity and quality.

This paper provides a comprehensive analysis of diffusion model loss functions, examining various formulations developed over recent years. The authors unify these objectives under the variational lower bound framework, offering both theoretical connections and empirical evaluations. Their research reveals when different objectives perform differently and which factors drive these variations. The study also assesses how loss function choice affects specific outcomes like sample quality and likelihood estimation.

Recently, several methods have leveraged deep generative modeling to produce example-based explanations of image classifiers. In this paper, we bridge the gap between these and the classical explainability literature by proposing a probabilistic framework for example-based explanations.

This paper introduces a vector-based approach for modeling and optimizing electric vehicle motor designs, addressing limitations of traditional variable-specific. Using a variational autoencoder trained on motor geometries. The method produces results comparable to current techniques while offering direct import into simulation for validation.

This paper enhances deep learning approaches for electrical machine simulation and optimization, addressing the time and resource constraints of traditional finite element analysis. The authors propose improvements to make both training and optimization processes more robust and efficient, while providing comparison with previously VAEs methods.