Related Neural Networks Links
Learn Transfer Learning Neural Networks Tutorial, validate concepts with Transfer Learning Neural Networks MCQ Questions, and prepare interviews through Transfer Learning Neural Networks Interview Questions and Answers.
Neural Networks
15 Essential Q&A
Interview Prep
Transfer Learning — 15 Interview Questions
Pretrained backbones, linear probes, layer-wise learning rates, and when transfer hurts (negative transfer).
Colored left borders per card; green / amber / red difficulty chips.
Pretrained
Freeze
Fine-tune
Domain
1 Define transfer learning.Easy
Answer: Reuse knowledge from a source task (often large-scale pretraining) to improve a target task with less data, time, or compute than training from scratch.
2 Feature extraction vs fine-tuning.Easy
Answer: Feature extraction: freeze backbone, train only classifier head on your labels. Fine-tuning: update backbone (all or top layers) with usually smaller LR than head.
3 When should you freeze early layers?Medium
Answer: Small dataset or high risk of overfitting—early layers learn generic edges/textures; later layers are more task-specific. Freeze first, then optionally unfreeze with care.
4 Why is ImageNet pretraining common?Easy
Answer: Large diverse natural-image supervision produces general low-level and mid-level filters that transfer well to many vision tasks (medical, satellite with caveats).
5 Discriminative learning rates.Medium
Answer: Use lower η for early layers, higher η for head/new layers—preserves generic features while adapting top of network.
6 Linear probe vs fine-tune (evaluation).Medium
Answer: Linear probe: train only linear layer on frozen features—measures representation quality. Full fine-tune measures end-task performance with adaptation.
7 Domain shift / domain gap.Medium
Answer: Source and target data differ in distribution—transfer may degrade; may need more fine-tuning, domain adaptation, or pretraining closer to target domain.
8 What is negative transfer?Hard
Answer: Pretrained model does worse than random init for target—source task misaligned or harmful biases; sometimes train-from-scratch wins on huge target data.
9 Transfer in NLP (BERT-style).Easy
Answer: Pretrain on large text with MLM/next-sentence; fine-tune whole model on downstream task with task-specific head—dominant paradigm before instruction-tuning era.
10 Replace only the classification head—when enough?Easy
Answer: When new task matches pretraining modality and semantics (e.g. 1000-way → N-way objects) and data is limited—fast and strong baseline.
11 Data augmentation with fine-tuning.Easy
Answer: Still critical—reduces overfitting on small target sets; match augmentations to target domain (e.g. no extreme color jitter on medical).
12 Catastrophic forgetting (mention in transfer context).Hard
Answer: Fine-tuning can erase source capabilities—relevant in continual learning; mitigated by small LR, elastic weight consolidation, or multi-task training.
13 Self-supervised pretraining then fine-tune.Medium
Answer: Pretrain without labels (contrastive, MAE, etc.), then supervised fine-tune—scales when labeled data is scarce but unlabeled data is plentiful.
14 More target data → train from scratch?Medium
Answer: With very large in-domain labeled data, gap shrinks; still often faster to start from pretrained weights. Empirical per task.
15 Interview one-liner: why use transfer learning?Easy
Answer: Better accuracy faster with less labeled data by reusing generic features learned at scale—standard for vision and NLP.
Mention freeze → unfreeze and smaller LR on backbone.
Quick review checklist
- Feature extraction vs fine-tuning; when to freeze.
- Discriminative LR; domain shift; negative transfer.
- Linear probe; NLP BERT pattern; self-supervised pretrain.