How do transfer learning and multi-task learning differ, with NLP examples?

• A. Transfer Learning Adapts A Model Trained On One Task To Another; Multi-Task Learning Trains On Several Tasks Simultaneously
• B. They Are The Same Concept
• C. Multi-Task Learning Uses Transfer Learning Only
• D. Transfer Learning Only Applies To Images

Answer: (B)

Transfer learning and multi-task learning are two different ways to share knowledge across NLP tasks. In transfer learning, you first train a model on a large source task or domain and then adapt it to a different target task, usually by fine-tuning the model on task-specific data. The idea is that the model learns general language patterns during pretraining or on the source task, and those patterns help when facing the target task with limited data. For example, you might start with a language model pre-trained on billions of words and fine-tune it to classify reviews as positive or negative.

Multi-task learning, by contrast, trains a single model to perform several tasks at the same time, sharing the same underlying representations. The model has a shared encoder and separate output heads for each task, and the training objective combines the losses from all tasks. In NLP, you could train a model to do part-of-speech tagging, named entity recognition, and dependency parsing together, so the shared representations learned from all tasks help each task perform better.

The key difference is where and how knowledge is transferred: transfer learning moves knowledge from one task or domain to another after separate training, while multi-task learning jointly learns multiple tasks to build shared, more general representations. Both ideas can be combined—pretraining on multiple tasks and then fine-tuning on a target task, or training on several NLP tasks and then specializing—depending on data availability and goals.