Transfer learning

📖 Full definition

Instead of starting from scratch, you take a model already trained on a large generalist dataset (e.g. ImageNet) and fine-tune it on a specific dataset. This cuts training time and cost by 100x and works better when you have little specific data.

💡 Concrete example

Look Like Me uses transfer learning: a CNN pre-trained on 10M faces is fine-tuned on family pairs to detect resemblances.

🔄 Synonyms

fine-tuning

🔗 See also

• CNN (Convolutional Neural Network) →
• Deep learning →
• ArcFace →