Essay

Analyzing Mismatch Hypotheses via Training and Training Dev Subsets

Question: If both your training and training dev sets contain a specific subset of data representing a mismatch hypothesis (such as car audio), explain how comparing performance on these subsets helps validate the decision to acquire more of that specific subset data.

Sample answer: By double-checking performance on the car audio subset in both sets, we can look for a performance gap. If the model performs well on the car audio in the training set but performs poorly on the car audio in the training dev set, this discrepancy validates the hypothesis that the model needs more car-recorded data to generalize properly. Thus, it justifies the effort to collect more car audio data.

Key points:

  • Identify and isolate the subset of interest (e.g., car audio) in both training and training dev sets.
  • Compare performance on this subset across both sets to check for a generalization gap.
  • Validate that more subset data is needed if the system performs well on the training subset but poorly on the training dev subset.

Rubric: The response must describe comparing subset performance on both the training and training dev sets, and explain that a high training subset performance coupled with low training dev subset performance validates the hypothesis that acquiring more of that data will help.

0

1

Updated 2026-05-26

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Related