引導式測驗 · 硬 等級

Echo State 網路和油藏計算測驗

檢查固定儲層動態、經過訓練的讀數、衰落記憶、光譜半徑以及儲層預測背後的計算。

問題 1 的 8

在迴聲狀態網路中,實際訓練了哪些權重?

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. 在迴聲狀態網路中,實際訓練了哪些權重?
    • 僅輸出(讀出)權重
    • 水庫內的所有循環重量
    • 僅輸入權重
    • 從未訓練過重量
  2. 哪個條件表達兩次相同輸入歷史下的 echo state 性質?
    • Their recurrent weights must all be trained by backpropagation
    • Their reservoir states eventually become independent of different initial states
    • Their output weights must have spectral radius exactly one
    • Their hidden states must equal the raw input at every step
  3. 為什麼訓練 ESN 比訓練標準 LSTM 快很多?
    • ESN 根本不使用神經元
    • ESN 僅在量子計算機上運行
    • ESN 僅透過求解線性讀數來避免時間反向傳播
    • ESN 不需要輸入數據
  4. 在調整 echo-state 儲備時,光譜半徑通常如何使用?
    • It directly sets the number of readout classes
    • It guarantees fading memory for every recurrent matrix
    • It measures the number of trainable reservoir weights
    • It is a recurrent-feedback scale heuristic for balancing rich dynamics and stability, not a complete guarantee of the echo state property
  5. 液體狀態機是一種尖峰神經元形式的儲存器計算,又是由誰引入的?
    • 沃夫岡·馬斯
    • 傑弗裡·辛頓
    • 吳恩達
    • 約書亞·本吉奧
  6. What does the fixed reservoir compute before the ESN readout produces an output?
    • A normalized probability over all possible weights
    • A high-dimensional dynamic state representation of recent inputs
    • A learned transition table for every future sequence
    • A separate trained policy for each output class
  7. What method is commonly used to fit an Echo State Network’s readout weights?
    • Backpropagation through all fixed recurrent weights
    • Dynamic routing by agreement
    • Linear least-squares regression, often with ridge regularization
    • A softmax over reservoir eigenvalues
  8. What does the spectral radius of a recurrent weight matrix mean?
    • The sum of all matrix entries
    • The number of nonzero output weights
    • The average magnitude of the input sequence
    • The largest absolute value among its eigenvalues