The Data-Driven Approach

We want to write some algorithm for classifying images into distinct categories, which scales much more naturally to all the variety of objects in the world. So the insight that makes this all work is the idea of the data-driven approach. Rather than sitting down and writing hand-specified rules to try to craft exactly what is a cat or a fish, instead we'll collect a large dataset of many, many cats/airplanes/deer/objects.

Once we get this dataset, we'll train this machine learning classifier which ingests all this data, then spits out a model which summarizes the knowledge of how to recognize these different object categories. Finally, we'll use this training model and apply it on new images to test its recognition of these objects, such as cats and dogs.

The Data Driven Approach:

• Relies on collecting a dataset of images and labels
• Uses ML algorithms to train a classifier
• Evaluates a classifier on new images