Superposition and Polysemanticity Quiz

Answers and explanations appear as you take the quiz.

1. What does 'superposition' refer to in neural networks?
   • Representing more features than there are neurons by overlapping their directions
   • Stacking multiple models on top of each other
   • Running the model twice and averaging
   • Adding noise to the inputs during training
2. What condition makes superposition work well despite feature interference?
   • Features being densely active all the time
   • Features being sparse, so few are active simultaneously
   • Using only positive activations
   • Having more neurons than features
3. What is a 'polysemantic' neuron?
   • A neuron that never activates
   • A neuron used only in the output layer
   • A neuron that responds to many unrelated concepts
   • A neuron with no connections
4. Which Anthropic paper introduced controlled study of this phenomenon in 2022?
   • Attention Is All You Need
   • ImageNet Classification with Deep CNNs
   • Playing Atari with Deep RL
   • Toy Models of Superposition
5. If a layer must store more features than it has dimensions, what is geometrically unavoidable?
   • Some features must share overlapping, non-orthogonal directions
   • The features can all stay perfectly orthogonal
   • The network must crash
   • All features become identical
6. Why does superposition make interpretability harder?
   • It makes the model slower to run
   • Single neurons no longer map cleanly to single concepts
   • It removes all features from the model
   • It only affects the input layer
7. Why does sparse feature activity reduce interference when more features than dimensions are represented?
   • Sparse activity makes all feature directions orthogonal
   • It increases the number of neurons automatically
   • Fewer features are active at the same time, so overlapping directions collide less often
   • It removes the need for nonlinear computation
8. What does an antipodal pair of feature directions look like geometrically?
   • Two vectors at a right angle in separate dimensions
   • Two vectors with identical directions and magnitudes
   • A vector with no coordinates
   • Two directions that point opposite ways along the same line