A pre-trained model is a neural network that has been trained on a relevant general dataset.

This model can then be used for specialized use-cases after going through a transfer learning process.

Transfer learning

Transfer learning is the process of specializing a pre-trained model for a certain use-case.

mlai

Link to original

For example, ResNet is a pre-trained model for computer vision. You can use ResNet to train models for use-cases such as classifying dogs and cats or curly and straight hair.

This is used for efficiency: it speeds up the training and requires a smaller data set.

An analogy: imagine you need to teach a chess opening. Would it be easier to teach which person:

  • One that already knows Chess: has already played before, knows how pieces moves, knows the ultimate goal
  • One that has never seen Chess before

In this analogy, the former is a pre-trained model and the latter is training from scratch.

mlai