Find out why an approximate decoder missed a better-scoring output.
Case context: A handwriting recognition system receives an image x and assigns each candidate text string T a score Scorex(T) = P(T|x). Because checking every possible string is infeasible, the team uses an approximate decoding method to search for the best output. On one test image, the system returns T1. Later, an engineer verifies that another valid transcription T2 has a higher score, so Scorex(T2) > Scorex(T1).
Question: Why did the decoder return T1 instead of T2, and what does this tell you about the decoding method?
Sample answer: The decoder returned T1 because the approximate search procedure did not find the highest-scoring transcription. This means the decoding method is not guaranteed to recover the true maximizer of Scorex(T) even when a better-scoring candidate exists.
Key points:
- The approximate decoder missed the best-scoring transcription.
- Approximate search methods can return a good but not optimal output.
- The higher-scoring transcription was missed because of search limitations, not because the score was computed incorrectly.
Rubric: The response must explain that: 1. The approximate search method failed to identify T2 even though it had the larger score. 2. This shows the method does not guarantee finding the absolute best output under the score function.
0
1
Tags
Data Science
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Beam Search as an Approximate Search Method
Why is it usually impractical to score every possible output sequence and choose the best one exactly?
True or False: Approximate search methods used during scored inference are guaranteed to return the highest-scoring result.
Because checking every possible output is impractical, a scored inference system must use a(n) _____ method to locate a high-scoring result.
Why is exact search impractical in a scored sentence-generation system?
Approximate search methods in scored inference are guaranteed to return the candidate S with the highest value of Score_A(S).
Search target in scored decoding
Match each inference term to its role in a scored output system.
Order the steps a speech transcription system uses to convert a recorded message into text.
Meaning of a Conditional Score in Classification
If each of N positions can be filled by any one of 50,000 words, then the number of distinct sequences of length N is (50,000)^N.
Approximate Search and Score Maximization
Match each search challenge to the concept that best describes it.
Order the reasoning steps that explain why approximate search is needed in scored inference.
Why exhaustive search is impractical in scored inference
Find out why an approximate decoder missed a better-scoring output.
State the cost of approximate search in large-scale scoring problems.