Let's assume that we want to train a binary classifier that can differentiate between Apple and Windows laptops based on real-time video feed using computer vision during corporate training. From the video feed, the first phase of model development would be to detect different objects and identify which of the objects are laptops. Once done, we can move to the second phase of formulating rules that can differentiate between Apple and Windows laptops. 

Now, there are already well-trained, well-tested open source models that can deal with the first phase of this model training. Why not use them as a starting point and use the inference toward the second phase, which is to differentiate between Windows and Apple laptops? This will give us a jump start and the solution will be less error-prone as phase 1 is already well tested.