In transfer learning, when is negative transfer MOST likely?

The correct answer is B. Source and target domains are unrelated. Transfer learning assumes that knowledge learned from the source task is useful for the target task. When source and target domains share similar structure (e.g., ImageNet features transferring to medical imaging), features generalize well. But when domains are unrelated (e.g…

NVIDIA Certified Associate (NCA) Core AI Concepts

Question

Options

ATarget dataset is small
BSource and target domains are unrelated
CModel is pretrained
DFine-tuning is applied

How the community answered

(26 responses)

A
4% (1)
B
92% (24)
D
4% (1)

Explanation

Transfer learning assumes that knowledge learned from the source task is useful for the target task. When source and target domains share similar structure (e.g., ImageNet features transferring to medical imaging), features generalize well. But when domains are unrelated (e.g., using a model trained on text sentiment to initialize a physics simulation predictor), the source representations encode patterns that are irrelevant or actively misleading for the target, causing negative transfer - performance is worse than training from scratch. A small target dataset (A), using a pretrained model (C), and applying fine-tuning (D) are standard transfer learning practices that by themselves do not cause negative transfer; the domain mismatch is the determining factor.

Topics

#Transfer Learning#Negative Transfer#Machine Learning Concepts

Community Discussion

No community discussion yet for this question.

Full NCA-AIIO Practice