3 ms·
I'm curious whether others here are working with supervised dimensionality reduction for face embeddings, particularly using single-task or multi-task learning
by rahimnathwani 2y ago
I'm curious whether others here are working with supervised dimensionality reduction for face embeddings, particularly using single-task or multi-task learning approaches.
While clustering tends to perform better after dimensionality reduction, selecting the optimal dimensions depends heavily on your specific use case. This makes it more complex than simply applying PCA or t-SNE.
- woodson 2y agoIf you have labelled data, you can try linear discriminant analysis (LDA; also known as canonical discriminant analysis), which maximizes the between-class variance while minimizing the within-class variance to best separate different classes by projecting data onto a new space that maximizes class separability (for whichever classes help your specific use case).