Bandar Bin Hazza

Artificial Intelligence Specialization · Cohort 2026

How to Regularize Autoencoders for Real-World Engineering Data? A Systematic Study Across Data Regimes and Latent Bottlenecks

We benchmark ten autoencoder regularization techniques — Vanilla, L1, L2, Contractive, Sparse, VAE, IRMAE, LoRAE, RRAE, and Denoising — on four synthetic PDE datasets of increasing difficulty: a translating Gaussian bump, thin-plate deflection, Gray–Scott reaction–diffusion, and a multi-channel shallow-water wave. All techniques share one convolutional backbone, one reconstruction loss, and one training budget, so differences come only from the penalty. We evaluate with an FID-style metric and latent-interpolation quality across multiple seeds, plus a per-technique hyperparameter ablation that separates real gains from seed noise. The outcome is a practical guide to regularizer choice for data-scarce engineering problems.

Read the final report (PDF)