Project story
This project proposes a Stacked Vector-Quantized Variational Autoencoder for classification on a dataset of microscopic white blood cells.
A stack of vector-quantized variational autoencoders compresses the image space into a lower-dimensional embedding. The model first learns reconstruction from unlabeled data, then performs classification from the embedding space with a labeled dataset.
Experiments examined few-shot learning and the effect of providing additional masked data during training.
Inside the work
