Adapter Layers for Transfer Quiz

Answers and explanations appear as you take the quiz.

1. What happens to the original pretrained weights when training with adapters?
   • They are frozen while only the adapter parameters are trained
   • They are deleted
   • They are randomly reinitialized
   • They are doubled in size
2. What is the typical internal structure of a bottleneck adapter?
   • A single huge dense layer
   • A down-projection, a nonlinearity, then an up-projection, with a residual connection
   • Only an attention head
   • A convolutional image filter
3. Roughly what fraction of parameters do adapters typically train?
   • 100%
   • Exactly 50%
   • Often under 5%
   • None at all
4. Why are adapters convenient for serving many tasks?
   • They require a separate full model copy per task
   • They cannot be combined
   • They erase the base model after each task
   • You keep one base model and swap small per-task adapters
5. To which family of techniques do adapters belong?
   • Parameter-efficient fine-tuning (PEFT)
   • Data augmentation
   • Image compression
   • Tokenization schemes
6. What advantage does LoRA offer over a classic inserted adapter at inference time?
   • It cannot be trained
   • Its learned weights can be merged into the base matrices, avoiding extra per-layer latency
   • It requires retraining the whole model
   • It removes the residual connection permanently
7. What does a residual connection contribute to a classic bottleneck adapter?
   • A second full pretrained model copy for each task
   • A mandatory external retrieval database
   • A path that adds the adapter’s transformed representation back to the original hidden state
   • A guarantee that no added computation occurs
8. What is one benefit of stacking a domain adapter with a task adapter?
   • Merge all user data into the frozen base weights automatically
   • Avoid using any adapter modules
   • Convert a language model into a vision encoder
   • Reuse a domain specialization together with a separate task specialization