Mamba and Selective State Spaces Quiz

Answers and explanations appear as you take the quiz.

1. What is the main computational advantage of Mamba over a standard Transformer?
   • It scales linearly with sequence length instead of quadratically
   • It needs no training data
   • It only works on images
   • It removes the need for any hidden state
2. What does the word 'selective' refer to in Mamba?
   • It selects which GPU to run on
   • Its SSM parameters depend on the input token, so it can choose what to remember or forget
   • It selects only nouns and verbs
   • It picks a random subset of tokens to ignore
3. How does a state space model represent the history of a sequence?
   • By storing every previous token explicitly
   • By drawing an attention map over all pairs
   • By compressing history into a fixed-size hidden state updated step by step
   • It does not keep any history
4. Why can't Mamba use the fast convolution shortcut that earlier SSMs used?
   • Because convolutions are illegal on GPUs
   • Because it has no parameters
   • Because it processes tokens in random order
   • Because making parameters input-dependent breaks the time-invariance convolutions require
5. What technique does Mamba use to stay fast despite losing the convolution shortcut?
   • A hardware-aware parallel scan kept in fast GPU memory
   • Running entirely on the CPU
   • Caching all token pairs
   • Skipping most of the sequence
6. Who introduced Mamba and roughly when?
   • OpenAI in 2018
   • Albert Gu and Tri Dao in late 2023
   • Google in 2015
   • No one, it is hypothetical
7. Which Mamba state-space parameters are made functions of the current input?
   • Only the fixed transition matrix A
   • The output vocabulary and tokenizer
   • B, C, and the step size delta
   • The attention query and key heads
8. What do the variable B, C, and delta parameters let Mamba do?
   • Build a growing key-value cache for every token
   • Route each token to an expert network only
   • Replace the state update with full all-pairs attention
   • Choose how much information to let into the state and how much to retain based on the input