Transfer Learning
A technique where a model developed for one task is adapted for a related task, reducing development time but requiring governance of inherited biases.
Definition
Involves fine-tuning pre-trained models on new, task-specific data. While efficient, transfer learning carries forward any biases or security weaknesses from the base model. Governance requires evaluating both base and fine-tuned model biases, tracking base-model provenance and licensing, and documenting fine-tuning datasets and hyperparameters to ensure legal and ethical compliance.
Real-World Example
A chatbot team fine-tunes an open-source BERT model on customer-service transcripts. Governance mandates a bias audit comparing sentiment across customer demographics on the base and fine-tuned models. They document the base-model license to confirm commercial use rights and log all fine-tuning parameters for reproducibility.
.svg)
