Training is the term we use to describe the process of creating a new model from scratch or updating an existing one to improve its performance in some way. There are two high-level steps to training a model: annotation and experimentation.
Think of it like football: before a team is ready for the big game, they need to practice. They’ll want to gather together all their equipment and head to the practice field. Once there, they’ll run drills over and over again, building muscle memory and learning how to anticipate what they might see from the other team on game day.
In this example, the team is the model, their equipment are the annotations, and the drills they run are the experiments.
In the annotation phase, you’re gathering together as many examples as you can of the object you’re looking for in your media (images or video). Just like the equipment in our football example, these annotations (also called labels) provide the material for training. A team can’t run drills without their cleats or pads, and a model can’t train without the individual annotations that tell it what to look for in an image.
In the experimentation phase, the CrowdAI platform takes all the annotations you created for your images and kicks off dozens of experiments at the same time. Continuing the football example: this is the phase where you run your drill over and over again to build muscle memory. Once the team has drilled a particular play long enough, they can move on to the next drill, until they know the whole playbook for the upcoming game.