Artificial Intelligence By Example

Making the model profitable by using it for another problem

The food processing experiment on the conveyor belt turns out to work well enough with dataset type d₁ and the CNN model M to encourage generalization to another dataset, d₂, in the same company.

Transfer learning consists of going from M(d₁) to M(d₂) using the same CNN model M, with some limited, cost-effective additional training. Variations will appear, but they will be solved by shifting a few parameters and working on the input data following some basic dataset preparation rules:

• Overfitting: When the model fits the training data quickly with 100% accuracy, this may  or not be a problem. In the case of classifying holes on the conveyor, belt overfitting might not prove critical. ...