引導式測驗 · 硬 等級

用於特徵提取測驗的稀疏自動編碼器

透過金門大橋實驗、超完備潛伏、轉向和模型邊界,研究稀疏特徵字典和重建。

問題 1 的 8

哪種可解釋性困難促成了神經激活的稀疏字典學習?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. 哪種可解釋性困難促成了神經激活的稀疏字典學習?
    • 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. SAE 稀疏性限制對給定輸入有何促進作用?
    • 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. 在 Anthropic 的「Scaling Monosemanticity」工作中,著名的範例功能是什麼?
    • 金門大橋特色
    • 數字七的特徵
    • 僅適用於藍色的功能
    • 元音的特徵
  4. 超完備(overcomplete)的 SAE 字典是什麼意思?
    • 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. 在這個研究情境中,單義特徵標籤旨在描述什麼?
    • 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. 如何使用 SAE 來影響模型的行為?
    • 透過開啟或關閉某個功能來控制輸出
    • 透過從頭開始重新訓練整個模型
    • 透過刪除標記器
    • 透過將其轉換為圖像模型
  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