Catastrophic Forgetting Quiz

Answers and explanations appear as you take the quiz.

1. Which situation describes catastrophic forgetting in a neural network?
   • When learning a new task causes a model to lose previously learned abilities
   • When a model runs out of memory during inference
   • When training data is accidentally deleted
   • When a model refuses to answer a question
2. Why does catastrophic forgetting happen in neural networks?
   • Because each task gets its own separate network automatically
   • Because new gradient updates overwrite shared weights that encoded old skills
   • Because models cannot store more than one fact
   • Because attention layers delete data on purpose
3. What does Elastic Weight Consolidation (EWC) do to reduce forgetting?
   • Deletes the new task's data
   • Doubles the learning rate
   • Penalizes changes to weights important for previously learned tasks
   • Removes attention from the model
4. How does experience replay (rehearsal) combat forgetting?
   • By freezing the entire model permanently
   • By training only on the newest data
   • By reducing the number of layers
   • By interleaving stored old examples during new training
5. What does LoRA do to base-model weights when adding a task-specific adapter?
   • It freezes base weights and trains a low-rank adapter, which may reduce direct interference but does not guarantee retention
   • It rewrites every base weight for each new task
   • It stores no task-specific parameters
   • It guarantees zero forgetting on every task sequence
6. How does a retrieval-augmented approach sidestep catastrophic forgetting?
   • By deleting old knowledge to make room
   • By storing new knowledge in an external database instead of the weights
   • By training continuously on random noise
   • By disabling the model's memory entirely
7. How does Gradient Episodic Memory (GEM) constrain learning on a new task?
   • It removes old task examples from memory
   • It changes the model to a nearest-class-mean classifier
   • It adjusts the gradient so the first-order approximation does not increase loss on stored prior-task examples
   • It adds a KL penalty to every parameter regardless of task
8. How do RAG-Token and RAG-Sequence differ in the original RAG paper?
   • RAG-Token removes retrieval while RAG-Sequence removes generation
   • RAG-Sequence uses images while RAG-Token uses audio
   • RAG-Token trains without model parameters
   • RAG-Token marginalizes retrieved documents per output token, while RAG-Sequence uses one document for a whole sequence