Training adjusts a model using examples and feedback about how well its outputs match expected results. For neural networks, learning means finding weight values that minimize a loss function; gradient descent provides a way to update them. section 1.4 section 20.1.1
Training also learns useful representations: transformations of input data that make a task easier. A neural-network workflow can combine layers, a loss function, an optimizer, performance metrics, and a mini-batch training loop. section 1.4 section 3.2
Training examples can take different forms:
- Speech recognition uses audio paired with transcripts; image tagging uses pictures paired with labels. section 1.4
- Self-supervised training can reconstruct missing or corrupted input, such as denoising images or predicting the next word. section 20.1.1
