Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ

Câu đố phân kỳ KL

Tính toán độ phân kỳ KL theo bit và lý do về sự bất đối xứng, hỗ trợ không khớp, entropy chéo và hướng phân phối tham chiếu.

Đường dẫn hướng dẫn liên quanKl phân kỳ
Câu hỏi 1 của 8

In D_KL(p||q), which distribution weights the log-ratio terms?

Xem trước mọi câu hỏi8 câu hỏi

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

  1. In D_KL(p||q), which distribution weights the log-ratio terms?
    • p
    • q
    • Their unweighted average
    • The distribution with larger entropy
  2. Why is KL divergence not an ordinary symmetric distance?
    • It can only be negative
    • Swapping p and q generally changes the value
    • It has no logarithms
    • It is always exactly one
  3. For p(x)>0 and q(x)=0, what happens to the corresponding forward KL contribution?
    • It is always one bit
    • Nó trở thành số không
    • Nó phân kỳ đến vô cùng
    • It is omitted regardless of p
  4. In the fair-versus-biased coin example, why can one summand be negative while total KL remains nonnegative?
    • The log base makes negative terms invalid
    • Every term is required to be positive
    • The formula takes an absolute value after summing
    • Individual log ratios can be negative, though the weighted total is nonnegative
  5. What unit results from using base-two logarithms?
    • Nats
    • Bits
    • Meters
    • Percent points
  6. How does cross-entropy relate to entropy and forward KL?
    • It is always zero when q differs
    • It equals D_KL(Q||P) minus entropy
    • H(P,Q) = H(P) + D_KL(P||Q)
    • It is the reciprocal of KL
  7. In SciPy's entropy(pk, qk), what does supplying qk request?
    • A test of whether probabilities are independent
    • Only Shannon entropy of qk
    • A symmetric distance between both arrays
    • Relative entropy using pk relative to qk
  8. Which statement about continuous KL is accurate?
    • It integrates a log density ratio under P and is infinite if P puts positive mass where Q has zero density
    • It sums point probabilities at every real-valued outcome
    • It is always finite whenever both distributions have densities
    • It has no support condition for the compared distributions