Pattern Recognition with Autoencoders: Binary Ellipses¶
As a necessity of analyzing experimental data taken in turbulent flows, I have been working with various analytical and numerical tools for extracting important flow patterns. The most common one I use is Proper Orthogonal Decomposition (POD), also known as Principal Component Analysis (PCA) in Machine Learning. Besides dimensional reduction, POD/PCA works great for extracting abstract representations of so-called coherent structures in the flow field that fluctuate periodically in a way resembling harmonic motions. However, due to its linear nature, POD/PCA is not always capable or efficient at capturing non-linear flow phenomena that are sporadic and chaotic. Increasingly, neural-network based unsupervised (or self-supervised) methods such as autoencoder have been adapted to tackle these problems in the field of fluid dynamics.
Table of Content¶
Initiated |
Title |
Note |
|---|---|---|
05.2021 |
with 2 latent codes |
|
06.2021 |
with 3 latent codes |
|
06.2021 |
PCA Autoencoder (Part 3): Hierarchy treatment (without rotation) |
with 2 sub-networks |
06.2021 |
PCA Autoencoder (Part 4): Hierarchy treatment (with rotation) |
with 3 sub-networks |
06.2021 |
Eigendecomposition |