In simple words: More data helps, and pre-training can reduce the data needed for a specific task.
Many adjustable weights
Deep networks often have many weights. They need enough varied examples to learn useful patterns and avoid merely memorising a small dataset. Data quality, labels, diversity, and a clean test set matter as much as raw volume.
Reuse earlier learning
A pre-trained model has already learned general features from a larger dataset. Transfer learning adapts it to a smaller task, often using fewer new examples. It still needs evaluation on the new task and population.
See the flow
Instead of training a photo model from zero for damaged parcels, a team can adapt a model already trained on general images.
Remember: More data helps, and pre-training can reduce the data needed for a specific task.