Intro
During the development of a machine learning model, data is fed into it for the purpose of training—similar to how a person studies by revising a test example and cross-checking the answers. Models are given sets of data: an input and the result it should produce. Over time, the model may gradually become very good at solving that particular set of tests; it may even score 100% accuracy when validated. However, when presented with a new test and no answers, it might fail to reproduce any usable outputs. This is referred to as a generalization problem—or more correctly, the model has overfit the initial data it was trained on. Using the test-answer analogy, it would be similar to a student studying a particular paper without learning how to reason about the underlying topic. When faced with a new set of problems, they begin to struggle.
“Adapt, improvise, overcome.” — Bear Grylls
tbc...