‘Teaching’ AI when defects are a rarity is a huge problem.
In automotive manufacturing, where lean Six Sigma practices have been widely adopted, most OEMs and Tier One suppliers strive to have fewer than three to four defects per million parts. The rarity of these defects makes it challenging to have sufficient defect data to train visual inspection models.
Technologies to circumvent the small data problem
Can manufacturing make AI work with small data? In fact, recent advances in AI are making this possible. Manufacturers can use the following techniques and technologies to circumvent the small data problem to help their AI projects go live even with only dozens or fewer examples:
- Synthetic data generation
- Anomaly detection
- Transfer learning
- Self-supervised learning
- few-shot learning
- Hand-coded knowledge
1. Synthetic data generation
Create synthetic data in areas where real data collection is very hard.
2. Anomaly detection
AI is trained with OK data only, ie: no defect data is feeded to AI. And the AI has to identify only whether the test subject is deviating from the Ok data or not.
3. Transfer learning
Learn from related tasks where there is ample data and use that knowledge in the small data tasks.
4. Self-supervised learning
Similar to transfer-learning, but the knowledge is acquired by solving a “slightly different” task.
5. Few-shot learning
AI is trained with thousands of tasks, each having a small number of examples( for ex: 10 examples ). That means AI is learning large number of small data sets.
One-shot learning is a special case of few-shot learning where the number of examples per task is only one.
6. Hand-coded knowledge
Modern machine learning use data instead of human institutional knowledge. But if small data is only available, we should use human institutional knowledge after perfecting it.