Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Sparse Autoencoders fun adanwo isediwon Ẹya

Study sparse feature dictionaries and reconstruction, with the Golden Gate Bridge experiment, overcomplete latents, steering, and model boundaries.

Jẹmọ itọsọna awọn ọnaSparse Autoencoders Fun Isediwon Ẹya
Ibeere 1 ti 8

Which interpretability difficulty motivates sparse dictionary learning for neural activations?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Which interpretability difficulty motivates sparse dictionary learning for neural activations?
    • Every model activation is already a complete dictionary definition
    • Several learned signals can overlap in an activation space
    • The tokenizer must be permanently removed
    • All neural weights are stored as image pixels
  2. What does an SAE sparsity constraint encourage for a given input?
    • Every feature to activate equally on every input
    • A mandatory removal of its decoder
    • A guarantee that every feature has exactly one perfect human label
    • Only a limited subset of learned feature activations to be nonzero
  3. Ninu iṣẹ 'Scaling Monosemanticity'' Anthropic, kini ẹya apẹẹrẹ olokiki?
    • The Golden Gate Bridge ẹya-ara
    • Ẹya kan fun nọmba meje
    • Ẹya kan fun awọ buluu nikan
    • Ẹya kan fun awọn faweli
  4. What does an overcomplete SAE dictionary mean?
    • It must activate every feature for every token
    • It contains no reconstruction mapping
    • It has more learned feature directions than input activation dimensions
    • It always requires less storage than the input vector
  5. What does a monosemantic feature label aim to describe in this research context?
    • A guarantee that the feature never responds to related examples
    • A feature associated with one interpretable concept
    • A language model that has only one parameter
    • A token represented by exactly one raw byte
  6. Bawo ni a ṣe le lo awọn SAE lati ni agba ihuwasi awoṣe kan?
    • Nipa didi ẹya kan tan tabi pa lati darí awọn abajade
    • Nipa retraining gbogbo awoṣe lati ibere
    • Nipa piparẹ tokenizer
    • Nipa iyipada si awoṣe aworan
  7. What does an SAE’s reconstruction residual measure for an activation vector?
    • The difference between two unrelated document titles
    • The total number of words in the model vocabulary
    • Only the model’s hardware temperature
    • The difference between the original activation and its decoded reconstruction
  8. When an SAE is fitted to collected activations of a frozen language model, which weights need to be trained for that fitting?
    • Every original language-model weight must be updated
    • Only the tokenizer’s character list
    • The SAE’s own encoder and decoder weights
    • No parameters of the SAE can change