MLA-C01 Question 94
Select 2You are building a machine learning model to classify images of animals into different species. During the pre-training phase, you notice that the dataset contains more images of some species (e.g., dogs and cats) than others (e.g., exotic birds and reptiles). Which pre-training bias metric(s) would help you assess the fairness of the dataset for this classification task?
- A
Class Imbalance (CI)
- B
Difference in Proportions of Labels (DPL)
- C
Confusion Matrix
- D
Image Resolution Consistency Metric
- E
KL Divergence of Label Distributions
Show answer and explanation
Correct answers: A, B
Explanation
In pre-training, it is essential to identify and mitigate biases in the dataset to ensure fair and robust model performance. Class Imbalance (CI) and Difference in Proportions of Labels (DPL) are both widely used metrics for detecting uneven label distributions in numeric, text, or image data. These metrics help identify overrepresented or underrepresented categories, which could negatively impact the fairness and accuracy of the model.
- A. Correct.
Class Imbalance (CI) is a key metric to assess if certain classes are overrepresented or underrepresented in the dataset, which is directly relevant to the fairness of the dataset.
- B. Correct.
Difference in Proportions of Labels (DPL) measures the disparity in label distributions, making it suitable for identifying bias in datasets with uneven class representation.
- C. Incorrect.
A confusion matrix is a post-training evaluation tool used to assess model performance, not a pre-training bias metric.
- D. Incorrect.
Image Resolution Consistency Metric evaluates the consistency of image resolutions in a dataset and is not related to fairness or label distribution.
- E. Incorrect.
KL Divergence of Label Distributions is a potential method to measure differences between distributions but is not specifically a pre-training bias metric used in standard AWS-related workflows.