Defect detection with scarce labelled data
Cormorant needed to detect rare manufacturing defects from camera images, but had only a few hundred labelled examples of each defect class. Off-the-shelf models and standard augmentation didn't generalize.
Technological uncertainty (T661 line 242)
“We used machine learning to build a defect-detection model, which was challenging with limited data.”
“It was uncertain whether acceptable detection accuracy on defect classes with fewer than 300 labelled samples was achievable. Standard transfer learning and augmentation had been tried and produced high false-negative rates, and it was unknown whether a synthetic-data or few-shot approach could close the gap without introducing artifacts the model would learn.”
The weak version says 'ML is hard'. The strong version pins the uncertainty to a measurable condition (few-shot, <300 samples) and states what was already ruled out.
Technological advancement (T661 line 246)
“We built a model that detects defects with good accuracy.”
“We advanced our understanding of few-shot defect detection: a physics-informed synthetic-augmentation pipeline combined with a contrastive pre-training step reduced false negatives materially in controlled tests, while we established that naïve GAN-generated samples degraded precision — a result that redirected the approach.”
The strong version reports the knowledge advance and a specific negative result, not a product milestone.
What backs a claim like this
Each claimed element ties to a source that shows it. This is the traceability a review tests.
| Source | What it shows |
|---|---|
| Experiment logs & metrics | The systematic runs, hypotheses tested, and per-class accuracy over time |
| Notebooks | The augmentation and pre-training experiments, including the discarded GAN approach |
| Model-version history | Which changes moved the metric, tying effort to result |
| Commit history | The dated iteration behind the experiments |
The takeaway
Claim the experiments, not the model. The eligible work is the systematic investigation into what would work under your constraints — including the paths that didn't.
Draft yours from real evidence
SREDlog reads your own GitHub history, documents and records to draft a narrative like the strong one — grounded in what you actually did.
3-day free trial · No credit card required