In order to perform deep learning, a machine has to have a formalized mathematical representation of what “correct” looks like. This is what the model optimizes for in order to minimize. This means that there has to be a codifable loss function in order to point machine learning at a problem.
Source
- Huberman Lab’s podcast episode with Lex Friedman