Transfer Learning is the practice of taking a model trained on one task or domain and adapting it to another, dramatically reducing the data and compute needed
Transfer learning reuses knowledge from one task for another: start from a model pre-trained at scale, then adapt, fine-tune some layers, train adapters, or simply prompt, for your target task. The pre-trained representations carry general structure (language, vision features) so the new task needs orders of magnitude less data.
Transfer learning is the economic foundation of modern AI: nobody trains from scratch for each problem. Every LLM application, fine-tune, and few-shot prompt is transfer learning in action: the entire build-on-foundation-models paradigm rests on it.
Fine-tuning is the most common transfer technique: take pre-trained weights, continue training on task data. Transfer learning is the umbrella: also covering feature extraction, adapters, and even prompting as inference-time transfer.
Large-scale pre-training learns general structure, syntax, semantics, visual features, that most tasks share. Adaptation only needs to learn the task-specific delta, hence small data sufficing.
On domain gaps too wide for the foundation (radiology images on a web-photo model), conflicting objectives, or catastrophic forgetting when adaptation overwrites general capability. Domain-closer foundations and parameter-efficient methods mitigate.