Transfer Learning
Transfer Learning
What is transfer learning?
- Reuse learned knowledge for a new task
- Fast, without learning from scratch
- Fine-tuning is the best-known method
What transfer learning really changes is the starting line of an AI project. Meetings used to begin with 'How do we collect tens of thousands of data points?' Now the first question is 'Which pretrained model should we use?' With a starting point closer to 80 than 0, the data needed drops from tens of thousands to a few hundred or thousand records, and training time shrinks from weeks to hours. So a project’s success depends less on modeling skill than on the judgment to pick a pretrained model close to your task.
In practice, there are three main approaches. Feature extraction freezes the early weights and attaches only a new final classification layer. Full fine-tuning readjusts all weights at a small learning rate. Lightweight techniques such as LoRA leave the original untouched and train only a small adapter. With a few hundred samples, feature extraction usually works; with several thousand or more, use fine-tuning. For large language models, lightweight techniques that cut the trained parameters to under 1% are effectively the default. It is important to distinguish this from retrieval-augmented generation (RAG), which plays a different role. Fine-tuning teaches tone, output format, and judgment criteria, while feeding in frequently changing factual information is RAG’s job.
The most common gap in practice is the expectation that 'training on our data will always make the model smarter.' Running full fine-tuning on 3,000 internal documents often causes catastrophic forgetting, where the model loses the summarization and translation it did well before. We also often see overfitting: 200 defect images trained over and over hit 98% validation accuracy, then drop to the 60% range on the actual line. When the pretraining data differs greatly from yours (general photos → X-ray images), transfer helps little or even gets in the way. Record the original model’s performance before touching it and compare against the same validation set. For commercial models, check the contract first to see whether the license permits fine-tuning and redeployment.
John KimAuthor of “B2B Marketing Design” and CEO of Nanum Consulting
Hyerang ParkContent storytelling and AI-driven brand content specialistA long-standing machine learning concept that became standard practice once large pretrained models (ImageNet, BERT, GPT) arrived.
Consider a hypothetical case. Company A, an industrial valve manufacturer with 240 employees, had always inspected every valve by eye. Four inspectors checked 1,200 units a day, and by afternoon more scratches slipped through, pushing customer complaints up to 7 a month. The company considered AI inspection, but it had collected only 320 defect images.
- The vendor said, 'You need at least 30,000 images.' Mr. Kim, QA team lead, estimated that photography alone would take a year, so he changed course: take an image model pretrained on ImageNet and layer only the company’s own data on top.
- Park, a manager on the production engineering team, relabeled the existing 320 images into three classes (scratch, foreign matter, normal), shot 780 more under different lighting angles, and split them into 980 for training and 120 for validation.
- The first attempt froze all the early layers and trained only the final classification layer. In two weeks it reached 88% validation accuracy, and Foreman Lee, a 20-year veteran, conceded, 'It gets the same borderline cases wrong that people do.'
- After confirming that most of the remaining misclassifications came from glare off the plating, the team fine-tuned only the top three layers further at a learning rate of 1e-5. It took 4 hours on a single GPU and KRW 60,000 (≈ US$45) in cloud costs.
- Six weeks in, after connecting it to the line, the defect detection rate rose from 71% to 94%. Manual re-inspection volume fell from 1,200 items a day to 180, and two inspectors moved to process improvement work.