Projects Feb 2025 Solo

Genomaly

Generative AI for anomaly detection: GAN vs VAE comparison on MNIST, then VAE-based defect detection applied to pharmaceutical capsule quality control.

  • Python
  • PyTorch
  • GAN
  • VAE

Two phases: a structured GAN vs VAE comparison on benchmark data, then an industrial application of VAE reconstruction-error anomaly detection to pharmaceutical manufacturing.

GAN vs VAE

Both models were trained on MNIST Digits and MNIST Fashion (60K training / 10K test samples per dataset) under identical conditions.

GANVAE
Image qualitySharperBlurrier; bottleneck discards high-frequency detail
Training stabilityHarder; minimax game collapses without careful tuning of LR, batch size, and discriminator dropoutEasier; explicit ELBO objective
Latent spaceUnstructured; no reliable semantic navigationProbabilistic via KL divergence; smooth geometry supports interpolation and class sampling

Industrial anomaly detection: pharmaceutical capsules

The second phase applies VAE to a real inspection problem: detecting defective tablets before they reach patients. A single defective capsule in a production batch is both a regulatory risk and a safety risk.

Dataset: the Capsule subset of MVTecAD, an industrial inspection benchmark. 219 normal capsule images for training; a held-out test set includes both normal and anomalous (broken, cracked, discolored) capsules. Source images are 1000x1000; resized to 64x64 for training.

Approach: the VAE is trained exclusively on normal capsules. At inference, reconstruction error is the anomaly score. A defective capsule reconstructs poorly because the model has no prior for anomalous structure; the error exceeds the threshold and triggers a flag.

19/23

Anomalies detected

82% on test set

128

Latent dimension

Hyperparameters
Batch 8; lr 1e-3; latent dim 128; 20 epochs; BCE + KL loss

The trained model also demonstrated reconstruction-based correction: when given a broken capsule image, the decoder reconstructed it with the break visually filled in, drawing from the learned distribution of intact capsule structure. The anomaly is visible by comparing the input against the reconstruction.