General Questions About AI
What is an AI model?
An AI model is a piece of software that is being trained to perform a specific task, e.g. inspect defects.
Why is the quality of the data important to an AI model?
The model learns what's in the dataset. Hence, the model performance depends heavily on the quality of the data. If the difference between what's being classified as rejected and approved is very small the model will have a hard time making decisions, like when we as humans try to take decisions on inadequate information.
How do you train an AI model?
Starting from a dataset containing information about “the truth”, e.g. what should be approved or rejected in a quality inspection. The information is used during model training. This is similar to first-day-at-work for a human performing quality inspections that needs training to decide whether a product should be classified as approved or rejected according to company standards..
When the dataset has been quality assured, Gimics AI models are trained on powerful graphics cards specialized in fast calculations. The model is then evaluated against test data it has never seen before (not being trained on) and can then be used for autonomous visual inspections.
How fast is Gimic's AI engine?
Gimics AI engine can handle about 30 images per second.
How many examples of approved and rejected products are needed to train the AI model to make correct decisions?
It varies from case to case. If the defects vary a lot in terms of color, shape and position usually more training data is needed. The model is continuously improved when new defects occur in production.
How are Gimics models improved?
The model is tuned and trained on more data.
What key factors are crucial for the models performance in production?
The quality of the training data, that is defects and non defects are being trained accordingly.
Lighting and camera position are very close to the conditions that were used when recording the training data.
The inspection requirements are crystal clear and reflected in the training and test data.