PCA Memory Efficiency

[joshuahwu]

PCA implemented in features.pca() is current a memory-limiting step in the analysis pipeline. We use the fast randomized implementation in the fbpca package.

Here are some things we should implement.

• Downsample frames in dataset. Calculate PC loadings on that set. Project rest of data using PC loadings.
• Save PC loadings to file.
• Batch updating of PC loadings w/incremental PCA
• Create tests using diffsnorm to measure difference in PC scores using these approximations
• Visualize eigenpostures