Synthetic Data for Machine Learning Applications
summary
The talk explains using FEM simulations plus unsupervised methods to create labeled synthetic training data for pipeline anomaly risk models, and argues these synthetics boost performance only if the ‘synthetic gap’ is measured and mitigated. Practical checks include using t-SNE to inspect manifolds and training classifiers to distinguish real vs generated samples, and it warns that mixing synthetic and real data improperly across cross‑validation folds and test sets biases evaluation—so folds must be constructed carefully.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 2:00Rosen and pipeline inspection
- 4:07Pipeline workflow challenges
- 6:55Labeling costs in vision
- 10:16Synthetic data alternative
- 14:10Physics simulations for data
- 17:02Tech Data’s synthetic CV
- 20:44Controlling synthetic data quality
- 23:41Clustering synthetic data
- 27:31Avoiding model discrimination
- 30:06Domain adaptation with synthetic data
- 33:24Future trends in synthetic data
- 35:48Model validation with synthetic data
- 37:24Conclusion & Q&A
related talks
20mSynthetic Data: The new Super Power for AI
36mArchitecting Data: A Deep Dive Into the world of Synthetic Data
25mThe Importance of Synthetic Data in Data-Centric AI
47mWhy Synthetic Data is needed for AI Fairness and Explainability
30m🤯 No data? No problem! Synthetic data to the rescue
01h 00mThe Synthetic Data Advantage: How It Helps You to Get Both AI Privacy & Explainability
29mSynthetic Data in Vision AI: Theory and Practice
20mSynthetic Data for Privacy, AI Fairness and Democratizing Data Access