Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·23h ago
AI Research
arXiv CS.AI·May 27·5 min read
Diffuse to Detect: Generative Diffusion Models for Unsupervised IC Anomaly Detection
TL;DR
Detecting defects in integrated circuits (ICs) typically requires labeled training data, which is expensive and time-consuming to obtain. Researchers applied diffusion models—generative AI systems that learn patterns by gradually adding noise to images—to identify IC anomalies without labeled examples.
✦ Why It Matters
Engineers can deploy anomaly detection in IC manufacturing without expensive labeled defect datasets, reducing time-to-production.
Key Takeaways
- Utilize a Diffusion Transformer for unsupervised anomaly detection.
- Compress raw measurements with an autoencoder for efficiency.
- Calculate anomaly scores using noise-prediction errors.
- Achieve interpretable results through latent-space analysis.
How It Works
The framework begins by compressing raw test measurements with an autoencoder, which reduces dimensionality while preserving essential features. The compressed data is then reshaped into a structured token sequence, incorporating sinusoidal and device-specific embeddings to enhance context.
Anomaly detection is performed by predicting noise during mid-range diffusion timesteps, allowing the model to identify deviations from expected patterns without needing labeled data.