Skip to content

Machine learning

Eigenfaces and latent face generation

Probabilistic PCA versus beta-VAE

Timeline
Jan 2025 to Feb 2025
Affiliation
Worcester Polytechnic Institute
Variance retained
95%+
Novel samples
100

What it is

Trained probabilistic PCA and beta-VAE models on 576-dimensional face vectors for unsupervised identity representation and generation, then compared reconstructions and latent interpolations.

Problem

How much of a face survives a 16-, 32- or 64-dimensional latent, and which model interpolates more smoothly.

Approach

Reduced to latent spaces of 16, 32 and 64 dimensions with 95%+ variance retained, generated 100 novel faces by decoder sampling, and compared models through PCA-projected visualizations and 10x10 output grids.

Highlights

  1. VAE showed stronger feature continuity and expressiveness; decoder sampling cut training time 35% over dense baselines.