Berlin 2024 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 34: Poster: Machine Learning, Data Science, and Reservoir Computing
DY 34.7: Poster
Mittwoch, 20. März 2024, 15:00–18:00, Poster C
Squeezing Sand Grains through a Bottleneck: Can Deep Learning Find a Minimal Description for Granular Particles? — •Azhar Akhmetova1 and Matthias Schröter1,2 — 1Georg-August-Universität Göttingen, Göttingen, Germany — 2Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany
The optimal number of parameters required for reliably describing a specific class of shapes is an open question. We test the possibility of retrieving the minimum number of parameters needed using the autoencoder architecture. Autoencoders are neural networks consisting of two parts: the encoder that compresses input data into a lower-dimensional representation, creating an information bottleneck. Then, the decoder reconstructs the original input from this compact representation. We start by applying autoencoders to shapes generated with a known parameter count based on their Fourier descriptors. The focus is on testing if the autoencoder's bottleneck dimension can measure the required number of parameters. Another open question is how the resolution influences particle shape description. The final aim of the project is to apply autoencoders on X-ray tomography data of real-world particles.
Keywords: granular matter; neural network; autoencoder; shape