Semi-supervised learning of imaging data
​
April 23, 2024
Schematic of the rotationally-invariant semi-supervised variational autoencoder (ss-rVAE, left) and disentanglement of representations for SEM images of gold nanoparticles create a nanoparticle library (right).
Scientific Achievement
We demonstrate a powerful approach for the classification of imaging data by generalizing learning from a small, labeled subset to a larger, unlabeled dataset where features of interest have strong rotational and positional disorder.​
​
Significance and Impact
The analysis of large datasets where clusters of nanoparticles exhibit positional and rotational disorder is a foundational problem in imaging methods ranging from optical to electron and scanning probe microscopy.​
​
Research Details
-
Developed a semi-supervised ss-rVAE to classify imaging data with significant noise.
-
Demonstrated improved accuracy and robustness on synthetic datasets.
-
Applied the ss-rVAE to experimental SEM data, effectively classifying gold nanoparticles at arbitrary rotations.