Feedback Learning

The more you teach it, the smarter it gets.

Every piece of operator feedback matters. Report false calls and defects to help the AI learn what is acceptable and what needs attention.

Book a demo AI AOI software

One click

From a wrong call to a training example Reported at the review screen, during production. Nothing leaves the machine.

Reported at the review screen, during production. Nothing leaves the machine.

80%	Fewer false calls Against a rule-based baseline, and it keeps improving with every report.

Against a rule-based baseline, and it keeps improving with every report.

Continuous learning

Improve the model with each round of feedback. Update after reviewing a false call, or collect feedback during production and update later.

Update after reviewing a false call, or collect feedback during production and update later.

You provide feedback. The AI learns from it.

Reporting a reviewed component as No error and confirming the feedback.

STEP 01 · OPERATOR

Mark it OK or NG

Mark the component as OK or NG. Your feedback can apply to other components in the same group.

Training feedback switch: off, on, then off. Feedback is recorded in both states; only on uses it for model training.

STEP 02 · OPERATOR

Choose what the AI learns from

Turn training on to include this feedback in model updates. Leave it off to record the feedback without using it for training.

Starting AI Update, processing the feedback, and completing the model update.

STEP 03 · THE AI

Update the AI

Run AI Update to learn from the selected feedback and adjust inspection parameters automatically.

Running the next recorded inspection and reviewing the resulting component alerts.

STEP 04 · MACHINE

Re-inspect

Inspect the board again and review the updated results.

Same part number, different supplier.

A different supplier can mean a different body shape or marking, even for the same part number. Mark an acceptable variation as OK so the AI can learn to recognize it across the component group.

Variation is not a defect

A different marking font or a change in body colour between batches does not necessarily mean a defect. Your feedback teaches the AI which variations are acceptable, while lifted leads, solder bridges and missing components remain inspection targets.

An update will not touch what you set by hand.

Your settings stay

An update works around anything you set yourself.

Manual settings take priority

Where your setting and the update disagree, yours is kept.

The line comes first

An update waits for the line, not the other way around.

Undo an update

Not what you wanted? Put it back the way it was.

How we stand out from the alternatives.

The same false call, handled three ways.

Tune thresholds	BY HAND
Generic AI AOI	VENDOR RETRAINS AND SHIPS IT
DaoAI	FEEDBACK AND AI UPDATE

What actually changesThe accept band widens for everything judged by that parameterNothing on your machine. The request goes to the vendor and the deployed model stays as it isThe model learnsIt gains an example of what that component looks like when correctEffect on real defectsDefects inside the widened band are accepted tooUnknown until the next model arrives, and it was not trained on your defectStill caughtA defect near the same score is still outside what is now normalWhere the decision livesA number somebody changed, and the reason is not stored with itWith the vendor, in a dataset you do not holdIn your datasetA labelled example in the product’s dataset, with the image attachedScope of the changeEverything that parameter governs, whether you meant it or notWhatever their retraining changed, across every plant running that modelThe group you chooseOne position, a class, a part number, or a packageWho can do itWhoever has permission to change inspection parametersAn AI engineer, usually theirsAny operatorCalls can be reported without permission to change the programLong termThe threshold stays loose and the reason is not recordedYou wait on a release cycle, and your images may have to leave the plant to get thereFixed for goodThe model has one more example and does not call that component again

Four questions we get every time.

How many reports before it stops calling that component?

It depends on the item and how different the component looks. The machine tells you when a product has collected enough to be worth an update.

Does reporting a false call change the program?

No. A report adds an example. The program changes when someone with that permission runs an update, and that can be undone.

Does the line stop while it retrains?

No. An update waits for the line rather than the other way around.

Will an update overwrite our engineer’s thresholds?

No. Anything you set by hand is what the machine keeps.

Want more control? Adjust inspection settings yourself, from individual thresholds to the tools used on each component.

FULL DEMO

Feedback learning end to end: report a false call, enable training, run AI Update, inspect again

Send us a board and we will tell you which machine it needs.

Book a demo 2D equipment 3D equipment AI AOI software